Knowledge reuse and ontologies

We are concerned with the development of a methodology and computational support environments to aid in the construction of knowledge-based systems. Two central research aims are:
  1. To enable human experts to model and analyse their problem solving expertise; this expertise analysis should be suitable to feed into the conceptual knowledge system design process.
  2. To enable and support reuse of previously constructed knowledge bases.
Of key importance to tackle these research aims is the development of problem solving ontologies. That is, we want to design ontology which provides primitive vocabularies expressive enough to describe and justify problem solving behaviors while being meaningful to human experts.

Ontology plays an important role in knowledge sharing and reuse. While ontology originally is defined as a "systematic theory of existence" in philosophy, we define ontology as "a system of concepts/vocabulary used as primitives for building artificial systems" from knowledge base perspective.

We view problem solving expertise as the product of an on-going process in which a structure on knowledge emerges as adaptation to a history of interactions with the problem solving environment. Knowledge being processed comes from various sources such as domain theory, objects being reasoned about, workplace environment, and so on. The emerging structure allows for effective application of this knowledge in a problem solving situation. Expertise is thus tuned to the specific environment in which problem solving is carried out.

Because of the specificity of problem solving knowledge, its reuse is limited. To allow for reuse of expertise, a technique of "knowledge decompilation" is widely recognized as being useful. This technique decomposes expertise into several kinds of knowledge, making explicit and justifying the role this knowledge plays in the problem solving process. Understanding knowledge content is a fundamental issue to allow for knowledge reuse and sharing.

Roughly speaking, problem solving knowledge can be decomposed into task-dependent and domain-dependent portions. The former is called task knowledge and the latter domain knowledge. Furthermore, task knowledge is deeply related to the environment, called workplace, in which the problem solving takes place. Careful study of workplace is necessary for us to discuss task knowledge which is sensitive to it. All the three kinds of knowledge require their own ontologies to make them reusable and sharable.

1. Knowledge modeling and ontology

In contrast with the majority's line of ontology research, we argue that designing ontologies for knowledge reuse requires commitments to the problem solving context, in general, and the task, in particular. (See our philosophy behaind the reseserch) As a consequence, we design different types of interacting ontologies.

Task ontology is a system of vocabulary for describing problem solving structure of all the existing tasks domain-independently. It is obtained by analyzing problem solving processes of domain experts and task structures of real world problems. Task ontology, which consists of generic verbs, generic nouns, generic adjectives, etc., provides primitives in terms of which domain experts can describe problem solving context and makes it easy to put domain knowledge into problem solving context, since it provides them with abstract roles of various objects which could be instantiated to domain-specific operations and objects. Domain knowledge organized without paying attention to its usage is difficult to find out how to incorporate what portion of it into a specific problem solving process.

Object ontology is a system of necessary and sufficient vocabulary to describe a class of systems to be reasoned about (e.g., digital systems), in a manner suitable for a particular class of tasks (e.g., diagnosis); object ontologies support reuse within that same class of tasks.

Model ontology is a system of generic vocabulary to refer to statements about a class of systems to be reasoned about independent of the problem solving task; model ontologies support reuse across task environments.

Object ontology guidelines justifies ways to refine or change concepts and relationships in a model ontology to tune representations to the problem solving task and/or to non-functional system requirements.

Workplace ontology is a system of vocabulary to capture boundary issues of the workplace in which problem solving is performed; these boundary issues (e.g., reliability of resources, limitations imposed by organizational processes) have a recognized effect on the task or domain model. Object ontologies augmented with workplace ontologies support reuse of representations over different workplaces within the same class of tasks.

Reports: 1994 and 1995

Related research on WWW:

2. Knowledge acquisition for knowledge-based systems

Knowledge acquisition is known as a serious bottleneck in building knowledge-based systems, since it is difficult to elicit expertise from domain experts. Efficient systems for supporting knowledge acquisition are badly needed to overcome this difficulty. Two systems for knowledge acquisition have been developing in our lab. One is ISAK and the other is MULTIS.

ISAK presents a new architecture for design knowledge acquisition by integrating a learning system based on EBL and interview system into a unified system in which the learning system acts as a driver of the interview process. ISAK first tries to explain a design drawing using its domain theory and obtains macro-knowledge for design when it succeeds in explanation. Otherwise, it performs interview with a domain expert to elicit knowledge necessary for understanding it.

MULTIS: MULti-Task Interview System generates a problem solving engine for an unknown task through task analysis interview with a domain expert using two-level task ontology: one is at a knowledge level and the other is at a symbol level. The former is discussed above and contributes to making the interface between domain experts and the system friendly and the latter is composed of building blocks which enable the system to generate an executable code of a problem solving engine. The knowledge level ontology has been fully evaluated by the KEs in eight companies who are qualified in building scheduling ESs.

In MULTIS, a new concept called generic process is introduced. A generic process is a pair of a generic noun and a generic verb. Examples of generic processes are Pick up a schedule recipient, Select a schedule resource, Relax a constraint, and so on. The whole structure of a domain expert's problem solving process is represented in the form of a network of generic processes, which is called a generic process network, GPN for short. GPNs are domain-independent framework of problem solving engines and stored in a case base for reuse in the task analysis interview.

All the generic verbs have several Lisp codes called building blocks associated with them. Building blocks are selected according to the generic verbs identified by interview and connected each other using the connection information obtained from GPN. Then, an executable scheduling engine capable of backtracking is generated.

Reports: 1994 and 1995

Related research on WWW:

3. Methodology for building reusable knowledge bases

A project of developing a methodology for building reusable knowledge bases has been done taking a substation restoration operation task in electric power networks as a target task. This task includes identification of fault sections, making a target configuration of the substations which can be operated using live facilities, and planning of how to build the configuration from the current one. The main idea behind this project is to extend the MULTIS methodology to that for building reusable knowledge bases. Domain knowledge is divided into two categories, field where problem solving takes places(i.e. medicine, machinery, electric power industry, etc.) and target that is a concrete object in the field (i.e. human body, automobile, power plan, etc.) They discusses cases where reuse is to be made across different tasks, fields, and targets.

Among several characteristics of this research, one to note is we verified our methodology by implementing three prototype ESs for substation operation according to it based on an operational ES for an existing substation. We eventually showed most of the domain knowledge except structural knowledge specific to each substation can be organized so as to be reusable across the three tasks and each task-specific knowledge represented in terms of task ontology is also reusable across substations whose structure is considerably different from each other.

4. Qualitative modeling and ontology

4.1 Device modeling and ontology

A qualitative model is a set of fundamental domain theories representing behavior and function of a target system in terms of the qualitative calculus for model-based problem solving. The main issues concerning model building for model-based problem solvers are reusability of the knowledge and ease of its description. On the basis of device-centered modeling, a model of the whole system consists of local device models independent of global structures. One of the goals of this research project is to find primitives for representing reusable device models of several systems, i.e., to create device ontologies.

We have created a device ontology for fluid and heat systems. The device ontology is based on causal specifications and global knowledge derived from general properties of the physical entity. (See more details about mechanisms supporting the ontology.) We have built reusable models of pumps, valves, heat exchangers etc. Our models include both behavior and function. More details about representing function and behavior are described in the following paragraphs.

4.2 Function and behavior modeling

Although a lot of researchers have pointed out the significance of functional representation, the general relations between function and behavior is not fully understood yet. We consider the knowledge of each component in a system as consisting of two elements. One is a necessary and sufficient information for simulation of the component which we call behavior.The other is the interpretation of the behavior under a desirable state which the component is expected to achieve, which we call function. By identifying primitives necessary for the interpretation of the behavior in various domains, we can capture what function is and represent it by selection and combination of them.

We are developing FBRL, Function and Behavior Representation Language, a language for representing function and behavior with the primitives we identified. FBRL represents a functional model as Behavior + Functional Topping(FT) where FT is a set of information for interpretation of the behavior. It consists of five kinds of information as follows:

  1. Goal of the component
  2. Function type
  3. Focus on a certain class of objects
  4. Focus on a certain stream of an object
  5. Necessity of output objects.
As we aim at promoting reusability of the functional models erepresented by FBRL, we have discussed about meaning which we want to present, i.e. ontology of function and behavior first, and then designed FBRL. As a result FBRL is rich in terms of meaning.

For evaluating FBRL, we are designing an explanation generator. Referring to functional models represented by FBRL, it generates explanations including functional information about object system. One important note is that an FBRL model enables the generator selecting appropriate functional terms for explanation at an appropriate level of abstraction. Using this characteristic and other information represented by FBRL models, this framework enables various types of explanations.

Further details about our ontology of function/behavior, FBRL, etc., are available on here.


MizLab Homepage