The Ontology Spectrum & Semantic Models

1. Dr. Leo Obrst MITRE Information Semantics Group Information Discovery & Understanding Center for Innovative Computing & Informatics January 12 & 19, 2006 The Ontology Spectrum & Semantic Models

2. Abstract The Ontology Spectrum describes a range of semantic models of increasing expressiveness and complexity: taxonomy, thesaurus, conceptual model, and logical theory. This presentation initially describes the Ontology Spectrum and important distinctions related to semantic models, e.g., the distinction among term, concept, and real world referent; the distinction among syntax, semantics, and pragmatics; the distinction between intension and extension; and de facto distinctions that the ISO 11179 standard makes (as do many others): data objects, classification objects, terminology objects, meaning objects, and the relationships among these. Then the individual semantic model types are discussed: weak and strong taxonomies, thesaurus, and weak and strong ontologies (conceptual model and logical theory, respectively). Each of these are defined, exemplified, and discussed with respect to when a more expressive model is needed. If time permits, semantic integration and interoperability are discussed with respect to the models. Finally, a pointer to an expansion of the logical theory portion of the Ontology Spectrum is given: the Logic Spectrum, which describes the range of less to more expressive logics used for ontology and knowledge representation.

3. Agenda • Semantic Models: What & How to Decide? • Information Semantics • Tightness of Coupling & Semantic Explicitness • Ontology and Ontologies • The Ontology Spectrum • Preliminary Concepts • Taxonomies • Thesauri • Conceptual Models: Weak Ontologies • Logical Theories: Strong Ontologies • Semantic Integration, Mapping • What Do We Want the Future to Be? • More:

4. Information Semantics • Provide semantic representation (meaning) for our systems, our data, our documents, our agents • Focus on machines more closely interacting at human conceptual level • Spans Ontologies, Knowledge Representation, Semantic Web, Semantics in NLP, Knowledge Management • Linking notion is Ontologies (rich formal models) • Content is King or should be! • And the better the content…

5. Internet Semantic Brokers Agent Programming Peer-to-peer Web Services: SOAP Community Applets Application N-Tier Architecture EAI Same Intranet Enterprise Middleware Web Same Wide Area Network Client-Server Same Local Area Network Distributed Systems OOP Systems of Systems Same OS Same CPU Linking From Synchronous Interaction to Asynchronous Communication Same Programming Language Compiling Same Process Space 1 System: Small Set of Developers Tightness of Coupling & Semantic Explicitness Explicit, Loose Far Performance = k / Integration_Flexibility Rules, Modal Policies Semantic Mappings OWL-S RDF/S, OWL Semantics Explicitness Web Services: UDDI, WSDL XML, XML Schema Data Workflow Ontologies Conceptual Models Taxonomies Data WHouses, Marts Federated DBs Same DBMS Same Address Space Local Looseness of Coupling Implicit, TIGHT

6. Ontology & Ontologies 1 • An ontology defines the terms used to describe and represent an area of knowledge (subject matter) • An ontology also is the model (set of concepts) for the meaning of those terms • An ontology thus defines the vocabulary and the meaning of that vocabulary • Ontologies are used by people, databases, and applications that need to share domain information • Domain: a specific subject area or area of knowledge, like medicine, tool manufacturing, real estate, automobile repair, financial management, etc. • Ontologies include computer-usable definitions of basic concepts in the domain and the relationships among them • They encode domain knowledge (modular) • Knowledge that spans domains (composable) • Make knowledge available (reusable)

7. Ontology & Ontologies 2 • The term ontologyhas been used to describe models with different degrees of structure (Ontology Spectrum) • Less structure:Taxonomies (Semio/Convera taxonomies, Yahoo hierarchy, biological taxonomy, UNSPSC), Database Schemas (many) and metadata schemes (ICML, ebXML, WSDL) • More Structure:Thesauri (WordNet, CALL, DTIC), Conceptual Models (OO models, UML) • Most Structure:Logical Theories (Ontolingua, TOVE, CYC, Semantic Web) • Ontologies are usually expressed in a logic-based language • Enabling detailed, sound, meaningful distinctions to be made among the classes, properties, & relations • More expressive meaning but maintain “computability” • Using ontologies, tomorrow's applications can be "intelligent” • Work at the human conceptual level • Ontologies are usually developed using special tools that can model rich semantics

8. Ontology & Ontologies 3 • Ontologies are typically developed by a team with individuals of two types • Domain Experts: have the knowledge of a specfic domain • Modelers (ontologists): know how to formally model domains, spanning domains, semantic properties, relations • On-going research investigates semi-automation of ontology development • State-of-art for next 100 years will be semi-automation • Humans have rich semantic models & understanding, machines poor so far • Want our machines to interact more closely at human concept level • The more & richer the knowledge sources developed & used, the easier it gets (bootstrapping, learning) • Rigorous ontology development methodologies evolving (e.g., Methontology), today’s practice is set of principles/processes • Tools are being developed that apply formal ontology analysis techniques to assist KR-naïve domain experts in building ontologies (OntoClean)

9. Ontology Spectrum: One View strong semantics Modal Logic First Order Logic Logical Theory Is Disjoint Subclass of with transitivity property Description Logic DAML+OIL, OWL From less to more expressive UML Conceptual Model Is Subclass of Semantic Interoperability RDF/S XTM Extended ER Thesaurus Has Narrower Meaning Than ER Structural Interoperability DB Schemas, XML Schema Taxonomy Is Sub-Classification of Relational Model, XML Syntactic Interoperability weak semantics

10. Problem: General Semantic Expressivity: High Problem: Local Semantic Expressivity: Low Problem: Very General Semantic Expressivity: Very High Problem: General Semantic Expressivity: Medium Ontology Spectrum: One View strong semantics Modal Logic First Order Logic Logical Theory Is Disjoint Subclass of with transitivity property Description Logic DAML+OIL, OWL From less to more expressive UML Conceptual Model Is Subclass of Semantic Interoperability RDF/S XTM Extended ER Thesaurus Has Narrower Meaning Than ER Structural Interoperability DB Schemas, XML Schema Taxonomy Is Sub-Classification of Relational Model, XML Syntactic Interoperability weak semantics

11. Triangle of Signification Intension <Joe_ Montana > Concepts Semantics: Meaning Reference/ Denotation Sense Real (& Possible) World Referents Terms “Joe” + “Montana” Syntax: Symbols Pragmatics: Use Extension

12. Concept Relations Term Relations Subclass of Narrower than Synonym Term vs. Concept • Term (terminology): • Natural language words or phrases that act as indices to the underlying meaning, i.e., the concept (or composition of concepts) • The syntax (e.g., string) that stands in for or is used to indicate the semantics (meaning) • Concept: • A unit of semantics (meaning), the node (entity) or link (relation) in the mental or knowledge representation model Concept Vehicle Term “Vehicle” Concept Ground_Vehicle Concept Automobile Term “Automobile” Term “Car”

13. Data Objects Classification Objects Terminology Objects Meaning Objects XML DTD Data Schema XML Schema Thesaurus Ontology Keyword List Term (can be multi-lingual) Data Attribute Data Element Data Value Documents Conceptual Model Taxonomy Instance Value Attribute Property Relation Privileged TaxonomicRelation Namespace Class Concept Example: Metadata Registry/Repository – Contains Objects + Classification

14. Root Directed Acyclic Graph Node Directed Edge Directed Cyclic Graph Tree vs. Graph Tree

15. Taxonomy: Definition • Taxonomy: • A way of classifying or categorizing a set of things, i.e., a classification in the form of a hierarchy (tree) • IT Taxonomy: • The classification of information entities in the form of a hierarchy (tree), according to the presumed relationships of the real world entities which they represent • Therefore: A taxonomy is a semantic (term or concept) hierarchy in which information entities are related by either: • The subclassification of relation (weak taxonomies) or • The subclass ofrelation (strong taxonomies) for concepts or the narrower than relation (thesauri) for terms • Only the subclass/narrower than relation is a subsumption (generalization/specialization) relation • Subsumption (generalization/specialization) relation: the mathematical subset relation • Mathematically, strong taxonomies, thesauri, conceptual models, and logical theories are minimally Partially Ordered Sets (posets), i.e., they are ordered by the subset relation • They may be mathematically something stronger (conceptual models and logical theories)

16. Taxonomies: Weak Example: Your Folder/Directory Structure • No consistent semantics for parent-child relationship: arbitrary Subclassification Relation • NOT ageneralization / specializationtaxonomy Example: UNSPSC

17. Taxonomies: Strong • Consistent semantics for parent-child relationship: Narrower than (terms) or Subclass (concepts) Relation • A generalization/specialization taxonomy • For concepts: Each information entity is distinguished by a property of the entity that makes it unique as a subclass of its parent entity (a synonym for property is attribute or quality) • For terms: each child term implicitly refers to a concept which is the subset of the concept referred to by its parent term HAMMER Claw Ball Peen Sledge • What are the distinguishing properties between these three hammers? • Form (physical property) • Function (functional property) • “Purpose proposes property” (form follows function) – for human artifacts, at least

18. animate object agent person organization employee manager Subclass of Two Examples of Strong TaxonomiesMany representations of trees Simple HR Taxonomy Linnaeus Biological Taxonomy

19. Another, mostly strong Taxonomy: Dewey Decimal System

20. When is a Taxonomy enough? • Weak taxonomy: • When you want semantically arbitrary parent-child term or concept relations, when the subclassification relation is enough • I.e., sometimes you just want users to navigate down a hierarchy for your specific purposes, e.g, a quasi-menu system where you want them to see locally (low in the taxonomy) what you had already displayed high in the taxonomy • Application-oriented taxonomies are like this • Then, in general, you are using weak term relations because the nodes are not really meant to be concepts, but only words or phrases that will be significant to the user or you as a classification devise • Strong taxonomy: • When you really want to use the semantically consistent narrower-than (terms) or subclass (concepts) relation (a true subsumption or subset relation) • When you want to partition your general conceptual space • When you want individual conceptual buckets • Note: the subclass relation only applies to concepts; it is not equivalent (but is similar) to the narrower-than relation that applies to terms in thesauri • You need more than a taxonomy if you need to either: • Using narrower than relation: Define term synonyms and cross-references to other associated terms, or • Using subclass relation: Define properties, attributes and values, relations, constraints, rules, on concepts

21. Thesaurus: Definition • From ANSI INISO 239.19-1993, (Revision of 239.194980): • A thesaurus is a controlled vocabulary arranged in a known order and structured so that equivalence, homographic, hierarchical, and associative relationships among terms are displayed clearly and identified by standardized relationship indicators • The primary purposes of a thesaurus areto facilitate retrieval of documents and to achieve consistency in the indexing of written or otherwise recorded documents and other items • Four Term Semantic Relationships: • Equivalence: synonymous terms • Homographic: terms spelled the same • Hierarchical: a term which is broader or narrower than another term • Associative: related term • A consistent semantics for the hierarchical parent-child relationship: broader than, narrower than • This hierarchical ordering is a Subsumption (i.e., generalization/specialization) relation • Can view just the narrower-than subsumption hierarchy as a term taxonomy • Unlike Strong subclass-based Taxonomy, Conceptual Model, & Logical Theory: the relation is between Terms, NOT Concepts

22. Thesaural Term Relationships

23. Thesaurus vs. Ontology Controlled Vocabulary Ontology Terms: Metal working machinery, equipment and supplies, metal-cutting machinery, metal-turning equipment, metal-milling equipment, milling insert, turning insert, etc. Relations: use, used-for, broader-term, narrower-term, related-term Concepts Logical-Conceptual Semantics (Strong) Thesaurus Real (& Possible) World Referents Terms Term Semantics (Weak) • ‘Semantic’ Relations: • Equivalent = • Used For (Synonym) UF • Broader Term BT • Narrower Term NT • Related Term RT Logical Concepts Entities: Metal working machinery, equipment and supplies, metal-cutting machinery, metal-turning equipment, metal-milling equipment, milling insert, turning insert, etc. Relations: subclass-of; instance-of; part-of; has-geometry; performs, used-on;etc. Properties: geometry; material; length; operation; UN/SPSC-code; ISO-code; etc. Values: 1; 2; 3; “2.5 inches”; “85-degree-diamond”; “231716”; “boring”; “drilling”; etc. Axioms/Rules:If milling-insert(X) & operation(Y) & material(Z)=HG_Steel & performs(X, Y, Z), then has-geometry(X, 85-degree-diamond). • Semantic Relations: • Subclass Of • Part Of • Arbitrary Relations • Meta-Properties on Relations

24. Narrower than Related to Center For Army Lessons Learned (CALL) Thesaurus Example imagery aerial imagery infrared imagery radar imagery combat support equipment radar photography moving target indicators intelligence and electronic warfare equipment imaging systems imaging radar infrared imaging systems

25. When is a Thesaurus enough? • When you don’t need to define the concepts of your model, but only the terms that refer to those concepts, i.e., to at least partially index those concepts • Ok, what does that mean? • If you need an ordered list of terms and their synonyms and loose connections to other terms (cross-references) • Examples: • If you need to use term buckets (sets or subsets) to use for term expansion in a keyword-based search engine • If you need a term classification index for a registry/repository, to guarantee uniqueness of terms and synonyms within a Community of Interest or namespace that might point to/index a concept node • You need more than a thesaurus if you need to define properties, attributes and values, relations, constraints, rules, on concepts • You need either a conceptual model (weak ontology) or a logical theory (strong ontology)

26. Conceptual Models: Weak Ontologies • Many conceptual domains cannot be expressed adequately with a taxonomy (nor with a thesaurus, which models term relationships, as opposed to concept relationships) • Conceptual models seek to model a portion of a domain that a database must contain data for or a system (or, recently, enterprise) must perform work for, by providing users with the type of functionality they require in that domain • UML is paradigmatic modeling language • Drawbacks: • Models mostly used for documentation, required human semantic interpretation • Limited machine usability because cannot directly interpret semantically • Primary reason: there is no Logic that UML is based on • You need more than a Conceptual Model if you need machine-interpretability (more than machine-processing) • You need a logical theory (high-end ontology)

27. Conceptual Model: UML Example Human Resource ConceptualModel

28. Logical Theories: Strong Ontologies • Can be either Frame-based or Axiomatic • Frame-based: node-and-link structured in languages which hide the logical expressions, entity-centric, like object-oriented modeling, centering on the entity class, its attributes, properties, relations/associations, and constraints/rules • Axiomatic: axiom/rule-structured in languages which expose the logical expressions, non-entity-centric, so axioms that refer to entities (classes, instances, their attributes, properties, relations, constraint/rules) can be distributed

29. Language L Models M(L) Ontology Intended models IM(L) Logical Theories: More Formally Conceptualization C * N. Guarino. 1998. Formal ontology in information systems, pp. 3-15. In Formal Ontology in Information Systems, N. Guarino, ed., Amsterdam: IOS Press. Proceedings of the First International Conference (FOIS’98), June 6-8, Trent, Italy. p. 7

30. Conceptualization B: Buyer Conceptualization S: Seller Conceptualization B1: Technical Buyer Conceptualization S1: Manufacturer Seller Conceptualization B2: Non-Technical Buyer Conceptualization S1: Distributor Seller Language LB1 Language LS1 Language LB2 Language LS2 Models MB1(LB1) Models MB2(LB2) Models MS2(LS2) Models MS1(LS1) Ontology Intended models IMB1(LB1) Intended models IMB1(LB1) Intended models IMB2(LB2) Intended models IMB1(LB1) A More Complex Picture (from E-Commerce)

31. Axioms, Inference Rules, Theorems, Theory Theory (1) Theorems are licensed by a valid proof using inference rules such as Modus Ponens (2) Theorems proven to be true can be added back in, to be acted on subsequently like axioms by inference rules Theorems Axioms (3) Possible other theorems (as yet unproven) (4) Ever expanding theory

32. Axioms Inference Rules Theorems Class(Thing) Class(Person) Class(Parent) Class(Child) If SubClass(X, Y) then X is a subset of Y. This also means that if A is a member of Class(X), then A is a member of Class(Y) SubClass(Person, Thing) SubClass(Parent, Person) SubClass(Child, Person) ParentOf(Parent, Child) NameOf(Person, String) AgeOf(Person, Integer) If X is a member of Class (Parent) and Y is a member of Class(Child), then  (X Y) And-introduction: given P, Q, it is valid to infer P  Q. Or-introduction: given P, it is valid to infer P  Q. And-elimination: given P  Q, it is valid to infer P. Excluded middle: P P (i.e., either something is true or its negation is true) Modus Ponens: given P  Q, P, it is valid to infer Q If P  Q are true, then so is P  Q. If X is a member of Class(Parent), then X is a member of Class(Person). If X is a member of Class(Child), then X is a member of Class(Person). If X is a member of Class(Child), then NameOf(X, Y) and Y is a String. If Person(JohnSmith), then  ParentOf(JohnSmith, JohnSmith).

33. Ontology Representation Levels Language Meta-Level to Object-Level Ontology (General) Meta-Level to Object-Level Knowledge Base (Particular)

35. (implies (isa ?BATTALION InfantryBattalion) (thereExistExactly 1 ?COMPANY (and (isa ?COMPANY Company-UnitDesignation) (isa ?COMPANY WeaponsUnit-MilitarySpecialty) (subOrgs-Direct ?BATTALION ?COMPANY) (subOrgs-Command ?BATTALION ?COMPANY)))) CYC MELD Expression Example Ontology/KRExpressible as Language and Graph • In ontology and knowledge bases, nodes are predicate, rule, variable, constant symbols, hence graph-based indexing, viewing • Links are connections between these symbols: Semantic Net! isa ?BATTALION implies InfantryBattalion thereExistExactly 1 1 and ?COMPANY isa ?COMPANY What’s important is the logic! Company-UnitDesignation isa WeaponsUnit-MilitarySpecialty) subOrgs-Direct subOrgs-Command

36. Ontology: General Picture at Object Level But Also This! Most General Thing Upper Ontology (Generic Common Knowledge) Processes Locations Organizations Products/Services Middle Ontology (Domain-spanning Knowledge) Metal Parts Lower Ontology (individual domains) Art Supplies Lowest Ontology (sub-domains) E-commerce Area of Interest Mostly This Washers

37. Upper Ontological Distinctions 1 Focus here is on a few of the many possible upper ontological distinctions to be made • Descriptive vs. Revisionary: how one characterizes the ‘ontological stance’, i.e., what an ontological engineering product is or should be • Revisionary: every model construct (concept) is a temporal object, i.e., necessarily has temporal properties • Descriptive: model constructs are not necessarily temporal objects • Multiplicative vs. Reductionist: how one characterizes the kinds and number of concepts to be modeled • Multiplicative: Concepts can include anything that reality seems to require or any distinction that is useful to make • Reductionist: Concepts are reduced to the fewest primitives from which it is possible to generate complex reality

38. Upper Ontological Distinctions 2 • Universal vs. Particular: the kinds of entities that ontologies address (the ‘universe of discourse’(s) of the ontology) • Universals: generic entities, which can have instances; classes • Particulars: specific entities, which are instances and can have no instances themselves • Continuant vs. Occurrent • Continuant: An entity whose identity continues to be recognizable over some extended interval of time (Sowa, 2000) • Occurrent: An entity that does not have a stable identity during any interval of time (Sowa, 2000) • 3-dimensional (endurant) vs. 4-dimensional (perdurant) • 3D view/ Endurant: an object that goes through time (endures), with identity/essence-defining properties that perhaps depend on occurrent objects but are not essentially constituted by those occurrent objects • 4D view/ Perdurant: an object that persists (perdures) through spacetime by way of having different temporal parts at what would be different times

39. Upper Ontological Distinctions 3 • Part & Whole: Mereology, Topology, Mereotopology, the ‘part of’ relation • Mereology: parthood, what constitutes a ‘part’? • Topology: connectedness among objects, what constitutes ‘connected to’? • Mereotopology: the typical contemporary analysis of ‘part of’ says that the relation requires both the notion of part and the notion of connectedness; neither is sufficient alone to describe what we mean by saying that something is a part of another thing

40. Summary of Ontology Spectrum: Scope, KR Construct, Parent-Child Relation, Processing Capability Ontology Spectrum Processing Scope Parent-Child Relation KR Construct Machine-readable Concept Term Machine-processible Sub-classification of Machine-interpretable Taxonomy SubClass of Narrower Than Thesaurus Strong Taxonomy Ontology Disjoint SubClass of with Transitivity, etc. Conceptual Model (weak ontology) Weak Taxonomy Logical Theory (strong ontology)

41. What do we want the future to be? • 2100 A.D: models, models, models • There are no human-programmed programming languages • There are only Models Transformations, Compilations INFRASTRUCTURE Ontological Models Knowledge Models Belief Models Application Models Presentation Models Target Platform Models Executable Code

42. Contact Questions? lobrst@mitre.org

43. Ontology Spectrum strong semantics Logic Spectrum on Next Slide will cover this area Modal Logic First Order Logic Logical Theory Is Disjoint Subclass of with transitivity property Description Logic From less to more expressive DAML+OIL, OWL UML Conceptual Model Is Subclass of Semantic Interoperability RDF/S XTM Extended ER Thesaurus Has Narrower Meaning Than ER Structural Interoperability DB Schemas, XML Schema Taxonomy Is Sub-Classification of Relational Model, XML Syntactic Interoperability weak semantics

44. Logic Spectrum: Classical Logics: PL to HOL most expressive SOL + Complex Types + Higher-order Predicates (i.e., those that take one or more other predicates as arguments) Higher Order Logic (HOL) From less to more expressive Logics Second Order Logic (SOL) FOL + Quantifiers (, ) over Predicates Modal Predicate Logic (Quantified Modal Logic) FOL + Modal operators First-Order Logic (FOL): Predicate Logic, Predicate Calculus PL + Predicates + Functions + Individuals + Quantifiers (, ) over Individuals Logic Programming (Horn Clauses) Syntactic Restriction of FOL Decidable fragments of FOL: unary predicates (concepts) & binary relations (roles) [max 3 vars] Description Logics ModalPropositional Logic PL + Modal operators (, ): necessity/possibility, obligatory/permitted, future/past, etc. Axiomatic systems: K, D, T, B, S4, S5 Propositional Logic (PL) Substructural Logics: focus on structural rules Propositions (True/False) + Logical Connectives (, , , , ) less expressive

45. Logic Spectrum: Semantic Web Languages: Ontologies & Rules most expressive Higher Order Logic (HOL) From less to more expressive Logics Second Order Logic (SOL) Modal Predicate Logic (Quantified Modal Logic) SOL extensions First-Order Logic (FOL): Predicate Logic, Predicate Calculus OWL-FOL SWRL OWL + Horn-like Rules Logic Programming (Horn Clauses) OWL Full Almost FOL, but Classes as Instances goes to SOL OWL DL Mostly SHOIN(D): Close to the SHIQ and SHOQ Description Logics OWL Lite Almost SHIF(D) (technically, it’s a variant of SHIN(D) ModalPropositional Logic RDF/S Positive existential subset of FOL: no negation, universal quantification Propositional Logic (PL) Linear Logic: consume antecedents Substructural Logics: focus on structural rules RuleML less expressive Expressed syntactically in XML, requires binding to a logic, ranges over all logics

46. Logic Spectrum: Other KR Languages, Query Languages most expressive Higher Order Logic (HOL) From less to more expressive Logics Second Order Logic (SOL) Modal Predicate Logic (Quantified Modal Logic) SOL extensions First-Order Logic (FOL): Predicate Logic, Predicate Calculus Knowledge Interchange Format (KIF), Common Logic (CL, SCL) CycL Constraint Logic Programming languages Logic Programming (Horn Clauses) OWL-QL Open Knowledge Base Connectivity Language (OKBC) Description Logics Datalog RDQL SPARQL XQuery XPath ModalPropositional Logic SQL Propositional Logic (PL) Linear Logic: consume antecedents Substructural Logics: focus on structural rules less expressive