A Step Towards Ontological Engineering

(Translation of the paper presented at the 12^th National Conference on AI of JSAI, pp.24-31, June, 1998)

Riichiro Mizoguchi

ISIR, Osaka University, Japan

URL: http://www.ei.sanken.osaka-u.ac.jp/

Abstract: We discuss ontology and ontological engineering which is expected to play a role of foundation of so-called Content-Directed AI Research. I first answer questions about what is content-directed research followed by some definitions of an ontology. I next answer further questions about an ontology and describe the merits of ontology design. Finally, I present my idea about Ontological Engineering as well as its future.

1. Introduction

   Three major paradigm shifts have been occurring in AI research community: (1) from process-centered to Information-centered, (2) from computer-centered to human-centered, and (3) from form-centered to content-centered. While the first two are important, the third is expected to play a critical role in the Knowledge society in the coming 21^st century and is the discipline of so-called “Content-directed research” which I have advocated for years.

   In AI, formalism-oriented research has dominated as basic research extensively to date. In other words, basic research concerning “vessels” has been done paying little consideration on its content, while research on “Theory of content”, that is, research on “what and how to put in the vessel” has not been done to date. This is partly because the content largely depends on the respective domains of interest, which prevents to investigate general theories of “content”. And this in turn has made it difficult for AI to treat the real world problems which require deep understanding of objects specific to respective domains together with effective devices to enable their treatment.

   My colleagues and I have proposed “Ontological Engineering”[Mizoguchi 96, 97] and have practiced it to date. This paper describes what is Ontological Engineering.

   　

2. Basic questions

   2.1 What is content-oriented research?

   Content-oriented research is a great attempt of knowledge sharing with humans and computers. The dichotomy of general theories of a vessel which can be easily formalized and “domain-specific knowledge” which lacks generality has been widely accepted in AI community. An idea that there cannot be any theory about “Content” has been accepted. The objective of Content-oriented research is to get rid of such a misleading understanding about AI research and to provide effective theories and techniques enabling “Knowledge accumulation” and “Knowledge interoperability” which do play critical roles in the next generation knowledge processing.

   2.2 How is it possible to come up with general theories of content for such a variety of domain knowledge?

   It is true that each branch of sciences has established their own knowledge respectively. On the other hand, computer science has pursued domain-independent theories and techniques treating “information” and “data” obtained by abstracting things existing in all the domains. It is thanks to appropriate abstraction and careful consideration about them. Similarly, content-oriented research tries to exploit abstraction and decomposability of knowledge providing sophisticated guidelines and tools for understanding and model building of the target world(See 5.1).

   2.3 Why ontology instead of knowledge?

   It is true that knowledge is domain-dependent, and hence knowledge engineering which directly investigates such knowledge has been suffering from a rather serious difficulty caused by its specificity and diversity. However, ontology is different. In the ontology research we investigate knowledge in terms of its origin and elements from which knowledge is constructed. Hierarchical structure of concepts and decomposability of knowledge are exploited to deeply investigate primitives of knowledge as well as background theories of knowledge which enables us to avoid the difficulties knowledge engineering has faced with.

   2.4 Isn’t Ontological engineering another application-oriented research?

   While ontological engineering deals with domain-specific knowledge, it tries to establish theories and technology for “accumulating” knowledge within reasonable size of stratified domains utilizing ontologies. It is such a branch of AI research that investigates basic theories and technology to treat real world knowledge and is such an enterprise that denies the simple dichotomy of AI research: “Basic research” and “Application research”.

   2.5 Does Content-oriented research neglect formalism-oriented research?

   Although it has a plan to improve the unhealthy tendency of AI community in which formalism-oriented research has been overemphasized to date, it never neglects the importance of “form” or “formalization”. It rather considers formalization as an indispensable technique as mathematics in physics. A typical example is seen in the use of axioms employed in the definitions of concepts and relations in an ontology. The content-oriented research tries to keep a healthy balance between “form” and “content” using an ontology as a kernel concept.

   　

   　

3. What is an ontology?

There are many interpretations about what an ontology is in spite of the fact that it is understood to serve as a kernel theory and building blocks of content-oriented research. In fact, hot discussions are often done in many meetings on ontology. This section presents some of the definitions of an ontology.

3.1 Some definitions

1. In philosophy, it means theory of existence. It tries to explain what exist in the world and how the world is configured by introducing a system of critical categories to account things and their intrinsic relations.

   　

2. From AI point of view, an ontology is defined as “explicit specification of conceptualization” [Gruber] which is widely accepted in AI community. “Conceptualization” here should be interpreted as “intensional” rather than “extensional” conceptualization contrary to that defined in [Genesereth and Nilsson, 1987].

   　

3. From knowledge-based systems point of view, it is defined as “a theory(system) of concepts/ vocabulary used as building blocks of information processing systems” by R. Mizoguchi. Knowledge bases are necessary when one builds a model of problem solving in the world. In such a context of problem solving, ontologies are divided into two types: Task ontology for problem solving process and domain ontology for domain where the task is performed.

   　

4. A looser definition is given by H. Takeda: An ontology is “A consensus about a target world of interest for a specific purpose”. The heart of this definition is to define an ontology as loose as possible without losing its essentials.

   　

5. Another given by Gruber[Gruber]: Ontologies are agreements about shared conceptualizations. Shared conceptualizations include conceptual frameworks for modeling domain knowledge; content-specific protocols for communication among inter-operating agents; and agreements about the representation of particular domain theories. In the knowledge sharing context, ontologies are specified in the form of definitions of representational vocabulary. A very simple case would be a type hierarchy, specifying classes and their subsumption relationships. Relational database schemata also serve as　ontologies by specifying the relations that can exist in some shared database and the integrity constraints that must hold for them.

   　

6. A compositional definition is given as follows: An ontology consists of concepts with definitions, hierarchical organization of them(not mandatory), relations among them(more than is-a and part-of), axioms to formalize the definitions and relations.

3.2 Typical questions

1. How is an ontology different from a knowledge base?

   Let me cite a phrase found in the email archive of ontology:

   ---------------------------------------

   Date: Wed, 26 Feb 1997 12:49:09 -0800 (PST)

   From: Adam Farquhar axf@HPP.Stanford.EDU

   　

   ………

   　　

2. Does it express the consensus knowledge of a community of people?

   　

3. Do people use it as a reference of precisely defined terms?

   　

4. Does it express the consensus knowledge of a community of agents?

   　

5. Is the language used expressive enough for people to say what they want to say?

   　

6. Can it be reused for multiple problem solving episodes?

   　

7. Is it stable?

   　

8. Can it be used to solve a variety of different sorts of problems?

   　

9. Can it be used as a starting point to construct multiple (sorts of) applications including:

a new knowledge base, a database schema, an object-oriented program?

The stronger the 'yes' answer is to these questions, the more 'ontological' it is.

……………

--------------------------------------------------------------------------

The above opinion is based on that there is no clear boundary between ontology and knowledge. It is a reasonable understanding when we think of Cyc[Lenat, 95] whose upper part is definitely an ontology and the whole seems to be a knowledge base. But, the above opinion is somewhat misleading, since there is no clear definition of a knowledge base. My answer to the question is that we need to introduce a relativity when we understand an ontology. That is, if we think of a rule base of an expert system as an example of a knowledge base, an ontology is apparently different from the rule base because the rule base is based on the creator’s conceptualization of the target world which the ontology specifies and formalizes. And, the rule base never represents such a conceptualization. Thus, such a relative relation between the both clearly shows the essential difference of the two.

(2) How an ontology is different from the class hierarchy in object-oriented paradigm?

They are similar and further, the developmental methodologies of an ontology and an object hierarchy is also similar to each other in the upper stream. In the lower stream, however, the former concentrate on declarative aspects and the latter on performance-related aspects. Thus, the essential difference between the two lies in that the ontology research exploits declarative representation, while the OO paradigm is intrinsically procedural. In OO paradigm, the meaning of class, relations among classes, and methods are procedurally embedded and they are implicit. The ontology paradigm, on the other hand, descriptions are made declaratively in most cases to maintain formality and explicitness.

(3) What’s new? How is it different from taxonomy of concepts?

An ontology contains a taxonomy as its component as discussed in the above. So, it partially implies a taxonomy. In general, a new term is rarely totally new. Rather, it is usually coined by extending existing terms. The term “ontology” is not an exceptional case. It is a new term and concept including existing concepts such as “taxonomy”, “common vocabulary”, “upper model”, etc. by adding formality, richer relations, explicit representation of stuff usually left implicit.

(4)Are upper model and domain ontology compatible with each other?

An ontology should be use-specific in engineering settings. People believe that a general and neutral ontology is of useless. A typical example is task ontology[Mizoguchi, 92, 95a, 95b] where domain knowledge has to be organized so as to fit into the task model the task ontology specifies. In such cases, neutral domain ontology cannot apply to any problem without adjusting to the task structure. This discrepancy suggests us a potential difficulty which we might face with. People involved in upper ontology development advocate the importance of general and use-neutral ontology which is hard to be accepted by people who have to treat real world problems. Ideally, however, many use-specific ontology could find some of the essential concepts in use-independent ontology which should be inherited. We need more effort to harmonize the both activities believing we will come up with several possible upper ontologies one of which is really agreeable.

(5) What is the computational semantics of an ontology? Is it just a set of labels?

This is one of the most crucial points of the roles an ontology plays. Contrary to that an ontology sometimes looks just a set of labels, it has deeper computational semantics. I have proposed the following three levels of ontologies.

Level 1: A structured collection of terms. The most fundamental task in ontology development is articulation of the world of interest, that is, elicitation of concepts and identifying so-called is-a hierarchy among them. These are indispensable to things to be an ontology. Typical examples of ontologies at this level include topic hierarchies found in internet search engines and tags used for metadata description. Little definitions of the concepts are made.

Level 2: In addition to that at the level 1 ontology, we can add formal definitions to prevent unexpected interpretation of the concepts and necessary relations and constraints also formally defined as a set of axioms. Relations are much richer than those at the level one. Definitions are declarative and formal to enable computers to interpret. The interpretability of an ontology at this level enables computers to answer questions about the models built based on the ontology. Many of the ontology building efforts aim at those at this level.

Level 3: The ontology at this level is executable in the sense that models built based on the ontology run using modules provided by some of the abstract codes associated with concepts in the ontology. Thus, it can answer questions about runtime performance of the models. Typical examples of this type are found in task ontologies[Mizoguchi, 92][Breuker94][Seta96][Chandra 98].　

(6) What are the concrete merits an ontology can provide?

This question is also very important. The following is an enumeration of the merits we can enjoy from an ontology:

1. A common vocabulary.

   The description of the target world needs a vocabulary agreed among people involved. The fundamental role of an ontology contributes to it.

   　

2. Explication of what is left implicit.

   In all of the human activities, we find presuppositions/assumptions which are left implicit. Typical examples include definitions of common and basic terms, relations and constraints among them, and viewpoints for interpreting the phenomena and target structure common to the tasks they are usually engaged in. Any knowledge base built is based on a conceptualization possessed by the builder and is usually implicit. An ontology is an explication of the very implicit knowledge. Such an explicit representation of assumptions and conceptualization is more than a simple explication. Although it might be hard to be properly appreciated by people who have no experience in such representation, its contribution to knowledge reuse and sharing is more than expectation considering that the implicitness has been one of the crucial causes of preventing knowledge sharing and reuse.

   　

3. Data structure

   An ontology in a database is the conceptual schema. In this sense, an ontology provides us with a data structure appropriate for information description and exchange.

   　

4. Systematization of knowledge

   Knowledge systematization requires well-established vocabulary/concepts in terms of which people describe phenomena, theories and target things under consideration. An ontology thus contributes to providing backbone of systematization of knowledge. Actually, I have been involved in a project on Knowledge systematization of production knowledge under the umbrella of IMS: Intelligent Manufacturing Systems project[IMS 97][Ueda 97].

   　

5. Standardization

   The success of the modern industries has been achieved thanks to standardization of various components. We cannot avoid standardization in the successful knowledge processing research and activities in the real world. We do need something compatible to “bolts” and “nuts” in our community.

   　

6. Design rationale

   Typical examples to be explicated include intention of the designers of artifacts, that is, part of design rationale. An ontology contributes to explication of assumptions, implicit preconditions required by the problems to solve as well as the conceptualization of the target object which reflects those assumptions. In the case of diagnostic systems, fault classes diagnosed and range of the diagnostic inference, in the case of qualitative reasoning systems, classes of causal relations derived, and so on.

   　

7. Meta-model function

   A model is usually built in the computer as an abstraction of the real target. And, an ontology provides us with concepts and relations among them which are used as building blocks of the model. Thus, an ontology specifies the models to build by giving guidelines and constraints which should be satisfied. This function is viewed as that at the metalevel.

   　

8. Theory of content

The following is an enumeration of the merits of the systems built based on an ontology:

1. Reusability

   The explicit description about assumptions of the world the system is interested in contributes to making the system understandable and transparent, and hence its reusability increases. Further, the level 2 ontology enables us to build a guidance system for identifying reusable modules for specific purpose. Needless to say, we need more to attain satisfactory reusability of modules.

   　

2. Interoperability

   An ontology helps us specify functions of the system and communication protocol among systems, it thus increase the interoperability of the systems.

   　

3. Ease of implementation

   An explicit specification of the assumptions, functions, and constraints of the modules make it easier to implement the system.

   　

4. Intelligent tools

4. A brief survey of the research on ontology

4.1 Theory

N. Guarino and J. Sowa have been independently conducting research on theories of ontology. The both share the attitude towards philosophy. They both incorporate the results obtained in philosophy as principles to design the top-level ontology. Many of the practitioners have negative attitude to the top-level ontology under which they are required to put their ontology because they believe no use-independent ontology is useful. In the case of building an ontology for a large scale knowledge base, however, the validity of the knowledge base necessarily be justified in terms of wider rage of tasks, that is, it needs to show its generality rather than task-specific utility. Compliance with the principled top-level ontology provides a good justification. Thus, top-level ontology is important.

Sowa’s ontology is based on J. S. Peirce’s idea of the firstness which is defined without assuming any other things like human, iron, etc., the secondness which is defined depending on other things like wife, teacher, etc., and thirdness which provides an environment or context where the secondness works like family, school, etc. He introduces two important concepts, continuant and occurrent in addition to the three and obtains 12 top level categories by combining the seven primitive properties[Sowa 95, 98].

Guarino’s top level ontology is more extensively incorporates philosophical consideration. It is designed based on mereology(theory of parts), theory of identity, and theory of dependency. His ontology consists of two world: An ontology of Particulars such as things which exist in the world and Universals which include concepts we need when we describe Particulars[Guarino 97].

4.2 Machine readable dictionary: MRD

Development of machine-readable dictionaries has been done extensively in natural language processing community where ontology has also been discussed as the upper level model of the words/concepts structure. Typical examples include WordNet [Miller93]　[WordNet],　EDR[Yokoi, 95],　EuroWordNet[EWN], Generalized Upper Model[GUM]. Cyc is not a dictionary, but is a huge common sense knowledge base whose upper level structure is an ontology. One thing to note here is the activity done in the ad hoc committee in ANSI where they intend to design a reference ontology(RO) (http://ksl-web.stanford.edu/onto-std/) by aligning the upper level model of each of WordNet, Cyc and EDR. It is not easy to attain the goal, but it is of worth to try, I think.

4.3 Metadata, XML, Tag, and intelligent player

Another activity to note is that about standardization about metadata such as Dublin Core[DC 97], MCF: Meta Content Framework [MCF 97] and RDF: Resource Description Framework [RDF 97] in W3C. Dublin Core is the first de facto standard of the metadata descriptors. MCF was proposed to W3C as a candidate of the framework for metadata representation and extended and elaborated into RDF which is becoming as a standard. RDF introduces XML for its syntax.

XML[XML] is a simplified version of SGML which is a powerful markup language definition language and has been widely used in document description for years. XML is equipped with several powerful hyper-reference functions to fit the internet environment. The good things of XML is that the users can define their own tags which shows not only structure information of the document but also its semantic information for various uses of the document to enable semantic interoperation. For example, we can design an intelligent instructional player of a teaching material as an XML document by sharing a set of tags for explicating the roles of the portion of document and controlling the interpretation. It can “Play” the XML document adaptively to the performance of the learners. Further, we could formalize the set of tags to specify the performance of such instructional players. While tags in metadata description generally form a level one ontology, such tags can be those at the second level. Intelligent players with plug & play capability with shared XML tags is expected to be promising with Java implementation.

4.4 Developing methodology

We here present some of the developing methodologies and environments.

4.4.1 IDEF5[IDEF5]

KBSI: Knowledge Based Systems Inc. has developed a comprehensive methodology for developing an ontology, IDEF5, especially for enterprise modeling together with graphical and representation languages. IDEF5 tries to follow the normal and typical methodology and adopts the steps such as term extraction, term definition, relation among terms extraction followed by formalization of them. Documentation is well done and contains helpful guidelines to follow.

4.4.2 TOVE[Tove]

TOVE is also one for enterprise ontology like IDEF5. Its remarkable characteristics include:

1. Requirements to the ontology are formulated as “Competency questions” which the resulting model based on the ontology has to answer. Thus, competence of the ontology is formally specified.

   　

2. In TOVE, axioms are represented in Prolog and answer the competency questions.

The approach is thus very formal and follows the standard approach in software engineering. The typical questions include:

1. What activities have to be performed to achieve a goal?

   　

2. Given some activities at time points in the future, what characteristics do the resource and activities have at time points other than those given.

   　

3. What if a task is shifted backward(forward).

TOVE thus axiomatizes time and activity to be able to investigate many characteristics of the model through conceptual-level execution.

4.4.3 DODDLE, Aspect theory and CLEPE&AFM

There are several activities also in Japan. DODDLE(A Domain Ontology rapiD DeveLopment Environment) [Yamaguchi 97] is an environment for ontology development which has been used to build a legal ontology. Its characteristics is that WordNet is incorporated in the environment as the upper model to give the users guidelines for identifying and defining domain-specific concepts.

Aspect theory [Takeda 95] is a framework for combining several ontologies built based on different viewpoints. An ontology is called an Aspect and a set of logical operation among them are defined to form a theory. ASPECTROL is a language to implement the theory. The ontologies built by the above two environment are those at level 1 mentioned in 3.2.

CLEPE[Seta 97] is an environment for task ontology[Mizoguchi, 93], a level 3 ontology,

Development. It is designed assuming three kinds of users such as basic ontology authors, task ontology authors and task model authors. Models are built under the guide of task ontology to maintain their consistency. Such models are able to answer the competency questions like TOVE.

AFM: Activity-First Method[Mizoguchi, 95] is an environment for ontology building where task analysis is first done to elicit verbs which specify objects necessary for performing the task of interest. Nouns used as objects of the verbs are then extracted with the roles played by them. The building process begins by document analysis managing several intermediate products obtained during the course of ontology building to enable sophisticated support. The ontology built by AFM is of the level 2. AFM is going to be augmented to enable users compare multiple ontologies built by different persons to come up with a partial agreement of them and to build a maximally agreed ontology.

5. Ontological Engineering

We here try to answer several questions about what an ontological engineering is.

5.1 What do I mean by “An ontology is a theory of content”?

Observation of physics tells us mathematics have played a critical role in establishing it as a beautiful science. Mathematics provides us with:

1. the concepts of natural number, real number, continuous function, analytical function, etc. which are well-formalized and rigorously defined, and

   　

2. the classes of operations on them and relations among them such as four rules of arithmetic, derivation, group, linearity, orthogonality, etc.

These concepts successfully cover what physics needs to explain phenomena it is interested in, which enables physics to be a theoretical science. That is, physics has a firm theoretical foundation supporting its activity in dealing with “content”

Then, how about knowledge processing research? Does it have a similar theoretical foundation for “content-directed research”? Let’s take “Logic” as an example. Viewing logic as a theory of knowledge representation, it treats “content” as a predicate which takes a truth value, true or false neglecting all the other characteristics of the rich meaning. In return of the simplification, logic succeeds in getting generality and beauty of the theory. A predicate P in P(X) means it is “something” taking true or false about an object X, nothing other. It does NOT tell X is red, heavy, a non-animate, a human, a student, a husband, X moves, walks, etc. In short, logic gives us nothing about meaning about the world.

An ontology provides us with a set of concepts(vocabulary) used as predicates we need to describe the world knowledge. It specifies what meaning and constrains each concept has other than taking a truth value. For example, it reveals the four predicates tall(X), human(X), student(X), move(X) are very different from each other by specifying them as “tall” is an attribute without identity, “human” is a category X belongs to by its inherent property with identity, “student” is a role concept which does not represent X’s inherent property and is not persistent like the process “move”. The four predicates thus have their own properties and constraints associated with the categories they belong to. An ontology explicates such content-related information in the form of axioms, that is, an ontology provides us with a theory of content.

The remarkable progress of physics have been attained by describing and accumulating knowledge in various domains thanks to exploitation of mathematical concepts and formalism. On the other hand, knowledge processing researchers have never tried accumulation of knowledge by saying “content of each piece of knowledge is too domain-specific”. ^* Including the discussion about up to what abstraction level from the top we can accept as a common theory of content, taxonomy in an ontology contributes to construction of the theory of content and to knowledge accumulation. Recognizing the fact that it is unlikely to have a powerful theory with small amount of concepts and operations like physics, we need to make an effort to establish the theory of content with a longer term perspective.

* The only exception of this is Cyc project[Lenat, 1995].

5.2 What topics does ontology engineering cover?

At least the following topics should be investigated.

Basic topics

Metaphysics(Ontology)

Metaphysics of science

Knowledge representation

Ontology development

Common sense base

Methodology of ontology development

Description language, Methodology, Environment

Ontology comparison, alignment and unification

Ontology application

Knowledge sharing and reuse

Knowledge management

Corporate knowledge

EDI/CALS/STEP

Business process modeling

Systematization of domain knowledge

Internet information retrieval

Standardization

Evaluation

Media processing

Media ontology

Common ontology for knowledge media

Media integration

5.3 Does ontological engineering have a methodology?

While methodologies and environments for ontology development have been proposed as described above, we cannot say we have one for ontological engineering. It should be what we need to explore during the course of daily activities towards our goal. The following is just my personal view of a guideline.

1. Investigate even what you have thought as a matter of course.

   　

2. Uncover hidden assumptions

   　

3. Formulate and formalize the fundamental concepts

   　

4. Formulate and formalize relations among concepts

   　

5. Conceptualize the process of the system itself

   　

6. Formulate and formalize the system in terms of what you explicate

   　

7. Etc.

5.3.1 What are implicit assumptions?

We cannot state anything without implicit assumptions. A typical case includes the situation where we build a model of a target object. In such a case, we find

1. Viewpoints set according to the use of model

   　

2. Assumptions dependent on the viewpoints and

   　

3. Fundamental concepts and relations among them

In most of the cases, however, these kinds of assumptions are kept hidden and what is represented is only the model built in terms of the fundamental vocabulary. Even worse, the basic terms in the vocabulary are seldom defined explicitly and we are often surprised to find there is little consensus about the definitions of them, say, “behavior”, “function”, or “model”. Thus, a model is usually so incomplete as an explicit description.

5.3.2 Isn’t there any concrete example?

We sometimes find we do not know what a term, we think as a matter of course, exactly means. The following is my colleague, Kitamura’s and my experience in the research of “Ontology of fault”[Kitamura 98].

What is a cause of a fault? At the first glance, it seems to be a very trivial question. But, the reality is not so. Let me take a fault of a TV set. Imagine, you are told by the engineer you asked to fix it “Sir, this condenser was broken, so I replaced it with the new one.” Even if you do not understand the causal relation from the condenser to the symptom you initially found, you easily accept that the condenser is “the cause” of the fault. If the TV set was very new, however, you might start to suspect another(deeper) reason which caused to break the condenser. An excessive current might have flown through the condenser because of a short in somewhere or a lot of heat might have been radiated by a device near it, etc.

The above two ways of understanding causes of a fault are different from each other. The former is based on the na�“the whole(TV set)” must have been induced by “a fault” of its “part”, and the latter based on the more physical causal understanding. In fact, the former needs only subsumption relation among parts to trace the causal chain, the latter needs deeper causal relations in physical processes.