Ontology Summit An Ontology Framework NIST Gaithersburg MD

Ontology Summit: An Ontology Framework NIST, Gaithersburg, MD April 22 -23, 2007 • Michael Gruninger • University of Toronto, Canada • Leo Obrst • MITRE, Mc. Lean, VA, USA University of Toronto •

Agenda • A Brief Inclusive Characterization of Ontologies • A Common Set of Characteristics for Ontologies • Ontology Dimensions: – Semantic • Degree of Structure and Formality • Expressiveness of the Knowledge Representation Language • Representational granularity – Pragmatic • • Intended Use Role of Automated Reasoning Descriptive vs. Prescriptive Design Methodology • Conclusions 2

A Brief Inclusive Characterization of Ontologies (in Computer Science & Engineering) • Ontologies are used to support sharable and reusable representations of knowledge – An early definition of Ontology: “a specification of a conceptualization” (Gruber, 1994) • Nevertheless, the sheer range of current work in ontologies: – Including taxonomies, thesauri, topic maps, conceptual models, and formal ontologies specified in various logical languages – Raises the possibility of ontologies being developed without a common understanding of their definition, implementation and applications • Our objective: – To provide a framework that ensures that we can support diversity without divergence – So that we can maintain sharability and reusability among the different approaches to ontologies 3

A Common Set of Characteristics for Ontologies • An ontology includes: – A vocabulary together with a specification of – The meanings of the terms in the vocabulary • This specification includes: – Identification of the fundamental categories in the domain – Identification of the ways in which members of the categories are related to each other – Constraining the ways in which the relationships can be used 4

Ontology Dimensions: Semantic & Pragmatic • We propose a set of dimensions that can be used to distinguish among different approaches – Semantic – Pragmatic • Semantic Dimensions: These constrain how a given approach specifies the meaning of the terms 1) Degree of Structure and Formality 2) Expressiveness of the Knowledge Representation Language 3) Representational granularity • Pragmatic Dimensions: These cover the context in which the ontology is designed and used 1) Intended Use 2) Role of Automated Reasoning 3) Descriptive vs. Prescriptive 4) Design Methodology 5

Ontology Semantic Dimension: (1) Degree of Formality (Structure) 1. Related to but not the same as the expressive power required of a representation language used to specify the ontology 2. Informal: An ontology can be specified in English or some other natural language in a document 1. This is an informal ontology, although it can be rich, unambiguous, precise 3. More Formal: A taxonomy can be term or concept based 1. Term-based: A topic hierarchy from more general terms at the top of the hierarchy to more specific terms as one descends through the hierarchy; 2. Concept-based: A hierarchy of classes in which the necessary and distinguishing properties of classes and their subclasses are represented 3. But, even if formalized, these models are very simply structured, i. e. , structured with a subsumption relation (narrower_than or subclass) 4. Very Formal: An ontology for engineering equations 1. Would specify the formal semantics of its terms (such as quantity and unit of measure) in a language enabling precision & unambiguous expression 5. Degree of Structure: that by which the vocabulary in an ontology is constrained and can support computation 6

Ontology Semantic Dimension: (2) Expressiveness of the Knowledge Representation Language 1. This is a dimension partially dependent on the first 1. Since an informal ontology will be expressed only as a list of terms or enumerated definitions in a natural language such as English 2. Although not a characteristic in itself for a given content, the language or framework that the content is modeled in: 1. Greatly affects, and in fact 2. Is the primary constraint on what can be expressed by the content 3. Predicates and individuals/instances, for example cannot be expressed in propositional logic 4. Many of the notions termed “ontology” cannot be expressed formally 5. Similarly, relationships besides subclass or narrower-than cannot be expressed in a strict or strong (i. e. , at least partially formalized) taxonomy, although they can be in a less strict (informal) taxonomy 6. In general, only if the KR language is logic-based will the ontology being machine-interpretable 7

Ontology Semantic Dimension: (3) Representational Granularity: 1. An ontology may contain terms and limited interrelationship representation 1. Example: a simple taxonomy 2. Example: a very formal ontology expressed in Common Logic but which contains only 3 classes and 2 properties 2. Or it may contain much more detail including many restrictions concerning how terms can relate to each other 1. Example: a very detailed English description of biological classes and discriminating properties 2. Example: a very formal ontology expressed in KIF, CL, Cyc-L, or OWL+SWRL which contains thousands of classes, thousands of properties, thousands of rules, and billions of instances/individuals 3. Some quantifiable metrics may give indications of the representational granularity: 1. Examples: average subclass/subproperty depth, average density/bushiness, number of axioms, etc. 8

Ontology Pragmatic Dimension: (1) Intended Use, Application Focus 1. The intended use may be: 1. To share knowledge bases 2. To enable communication among software agents, 3. To help integrate disparate data sets 4. To represent a natural language vocabulary 5. To help provide knowledge-enhanced search 6. To provide a starting-point for building knowledge systems 7. To provide a conceptual framework for indexing content, etc. 8. The intended use often means that: 2. Typically, there is some application that is envisioned for which the ontology is being developed 1. Categorization: one might want to situate documents within a framing topic taxonomy that roughly characterizes the primary content of the document and helps one semantically loosely organize document collections 2. Search Enhancement: one might want to use a thesaurus, its synonyms and narrower-than terms, to enhance a search engine that can employ query term expansion, expanding the user’s text search terms to include synonyms or more specific terms and thus increasing the recall (i. e. , total set of relevant items) of retrieved documents 3. Enterprise Modeling, Question-Answering, Semantic Web Service Discovery 4. Semantic Search: one might to use “concept” based rather than “term” based search in specific domains or across domains 5. Complex Decision-Making, Intelligence Analysis: requiring machine reasoning 9

Ontology Pragmatic Dimension: (2) Role of Automated Reasoning 1. Automated reasoning can range from simple to complex 2. Simple automated reasoning can mean: 1. Machine semantic interpretability of the content, which only requires that the language that the content is modeled in is a logic 2. This is a principled or standards-based approach 3. Or it can mean that a special interpreter/inference engine has been constructed that knows how to interpret the content 1. This is an ad hoc and often proprietary approach 10

Ontology Pragmatic Dimension: (2) Role of Automated Reasoning (cont’d) 1. Simple automated reasoning: 1. The machine may be able to make inferences using the subclass relation by which properties defined at the parent class are inherited down to the children classes; this is the property of transitivity 2. More complex reasoning: 1. The machine may be able to take an arbitrary assertion in the KR language and classify it with respect to the taxonomic backbone of the ontology; e. g. , description logics perform classificational reasoning 3. Very complex automated reasoning: 1. The use of deductive rules, i. e. , inference rules or expressions that combine information from across the ontology 1. These characterize dependencies much like if-then-else statements in programming languages 2. Business rules that try to characterize things that have to hold in an enterprise but which can’t typically be expressed in relational databases or object models 3. A logic-based KR language needed: ‘validly concludes’ or ‘X is consistent with Y’ are not expressible generally in ad hoc implementations 2. Theorem-proving, theorem-generation, etc. 11

Ontology Pragmatic Dimension: (3) Descriptive vs. Prescriptive 1. Descriptive: 1. Does the content describe, i. e. , characterize the entities and their relationships as a user or an expert might characterize those objects? 2. Descriptive often takes a looser notion of characterization, perhaps allowing arbitrary objects into the model, which might not exist in the real world but which are significant conceptual items for the given user community 3. Potential partial synonym: Multiplicative, i. e. , concepts can include anything that reality seems to require or any distinction that is useful to make 2. Prescriptive: 1. Does the content prescribe, i. e. , mandate the way that those entities and their relationships are characterized? 2. Prescriptive often takes a stricter notion of characterization, stating that only objects which actually exist or that represent natural kinds or types of things in the real world should be represented in the content of the engineering model 3. Potential partial synonym: Reductionist, i. e. , concepts are reduced to the fewest primitives from which it is possible to generate complex reality 12

Ontology Pragmatic Dimension: (4) Design Methodology 1. Was there a methodology employed in the construction of the ontology? Possible ranges of methodology include: 2. Bottom-up: A bottom-up (sometimes called “empirical”) methodology places strong emphasis on: 1. Either solely analyzing the data sources so that the resulting ontology covers their semantics 2. Or enabling arbitrary persons to characterize their content as they personally see fit, using terminology or metadata and whatever structuring relations (or not) that they desire to use, with perhaps an auxiliary notion or assumption that in by doing so, patterns of characterizations may emerge or be preferred by a large group or community of persons 3. Can be conceptually profligate, tolerable of redundancy and partial overlap 3. Top-down: A top-down (sometimes called “rationalist”) methodology places strong emphasis on: 1. Developing the ontology using known notions about the world or domain 2. Independent of existing data sources whose semantics will be covered by the resulting ontology 3. Considering a range of questions that a domain expert might want to ask about the domain 4. Typically preferring a rigorous methodology focused on consistency, parsimony, etc. 13

Other Dimensions? • Semantic Focus (originally: Concept-based? ): – Term (0) – Concept (1) – Real World Referent (3) – Example: thesauri & some taxonomies: Term – Example: some conceptual models, some taxonomies, some ontologies: Concept – Example: some ontologies: RW Referent • Precision: subsumed by Granularity? • Scope: subsumed by composition of Application Foci? • Human-coded vs. Machine-generated • Range of values for the Ontology Types? – Scale? – Description only? – Both? 14

Conclusions • We want to avoid positing definition, allow definition to emerge • The framework is intended to offer dimensions for comparison – How our “ontologies” are alike, how they are different but comparable • For too long we have been embroiled in endless argumentation • Let’s cooperate, find correspondences, figure out how we can semantically interoperate • We’ll have a number of presentations & discussions to help bring the communities together • Goal: Converge! 15