KARATEKIT:
Tools for the Knowledge-Creating Company

Andreas Abecker*, Stuart Aitken+, Franz Schmalhofer*,
and Bidjan Tschaitschian*

*German Research Center for Artificial Intelligence (DFKI)

P.O. Box 2080, D-67608 Kaiserslautern, Germany
Andreas.Abecker@dfki.de

+Artificial Intelligence Applications Institute (AIAI)

University of Edinburgh, 80 South Bridge, Edinburgh EH1 1HN, Scotland
stuart@aiai.ed.ac.uk


Abstract. The companies of the future will live in an environment where markets are continuously shifting, technology proliferating, competitors multiplying, and products become obsolete overnight. Under such circumstances, knowledge management (in general) and knowledge creation (in particular) are becoming the most decisive business factors. Knowledge creation comprises the social sharing of vague, ambiguous and contradicting information, and thereby discovering novel concepts (Nonaka & Takeuchi, 1997). In the current paper, we describe how the KARAT and EKI Tools (KARATEKIT) can be jointly applied to support the knowledge-creation process. The KARAT system integrates techniques from knowledge acquisition, hypertext, text analysis, and groupware to support sharing of informal and tacit knowledge as well as its stepwise and gradual formalization. The EKI tool employs a user-programmable marker-passing mechanism that allows to reason beyond a fixed conceptual representation space. In this way the frame of reference can be extended and interesting new relations or novel concepts may be identified. The application of our system components is illustrated with respect to the historical knowledge-creation process at Honda in 1978, where the car concept "tall boy" was invented (i.e. a car being short for dealing with high density of traffic and being tall for maximizing passenger space).

Keywords: organisational learning, knowledge management, creativity

1 Motivation

In the currently emerging knowledge society, knowledge is seen as the most important success factor. Similar to the significance of technologies and mechanical machines during the industrial revolution, knowledge creation, knowledge management, and organizational innovations will play the pivotal role in the future's businesses (Drucker, 1993). The creation and acquisition of knowledge, as well as its efficient utilization will be the most decisive factors for maintaining or achieving the leading edge in successful markets. Under these circumstances, it will make perfect sense to talk about successful products as being the coordinated and reified knowledge of some business enterprise.

Despite of the very substantial successes of information technology in modern industry, this technology has not increased the productivity of modern businesses in the expected magnitude (Landauer, 1995). The disappointment about the overall benefits of current information technology in industry is clearly expressed by the catchy phrase that "information systems would show up everywhere in the world except in productivity statistics". And it is indeed true that the very practical problems of managing knowledge in some enterprise is not yet solved satisfactorily. Even the most innovative approaches primarily deal with knowledge conservation, distribution, and sharing (see, e.g. (Kühn & Abecker, 1997)), whereas knowledge-creation is mostly neglected.

In their seminal book about the "The Knowledge-Creating Company", Nonaka & Takeuchi (1995) have recently analyzed why the problems of knowledge creation and organizational learning are still not mastered with the highest possible competence in companies of the western hemisphere. More specifically, they inspected several success cases of Japanese companies, and they showed how these companies create the dynamics of innovation. Confronted with an environment where markets were shifting, technology proliferating, competitors multiplying and products becoming obsolete overnight, these companies have realized that the only certainty is uncertainty. Nonaka & Takeuchi's analysis was done from a management point of view without especially considering the question of IT support.

If we want to build upon Nonaka & Takeuchi's results, and also support knowledge-creating processes with IT, we must investigate how knowledge-based systems and AI can take into consideration needs for change and innovation. Looking at the current state-of-the-art in AI, we find mostly theories, models, and technologies which were developed with the objective of providing computer solutions for a relatively stable set of requirements and at least for a fixed frame of reference (see: Clancey, 1991). Model-based expert systems (Wielinga, Schreiber & Breuker, 1992), or formal ontologies (Gruber, 1993) are a good example for such a fixed frame of reference, which was designed for the purpose of knowledge sharing and reuse.

While such techniques build on stability, Nonaka & Takeuchi's results teach us to focus on change. In this paper, we discuss how specific tools from AI (namely our KARAT and EKI tool, resp.) can adjust their techniques towards this requirement. In the next section, we will briefly review Nonaka & Takeuchi's approach. We will use their suggested notion of knowledge, their distinction between tacit and explicit knowledge, as well as their process model for organizational knowledge creation. Our goal is to suggest how IT can contribute to Nonaka & Takeuchi's aims. For illustrating our ideas, in section 3, we will show how the historical example of developing a new-concept car, as it has been documented by Nonaka could be at least partially performed with the information technologies provided by the KARAT and EKI systems. In section 4 and 5, the two systems are then described in more detail. We conclude with section 6.

2 A Process Model for Knowledge Creation and Organisational Learning

Recently, (Choo, 1996) concisely summarized Nonaka & Takeuchi's comprehensive model of organizational knowledge creation:

" Knowledge creation is achieved through a recognition of the synergistic relationship between tacit and explicit knowledge in the organization, and through the design of social processes that create new knowledge by converting tacit knowledge into explicit knowledge. Tacit knowledge is personal knowledge that is hard to formalize or communicate to others. It consists of subjective know-how, insights, and intuitions that come to a person from having been immersed in an activity for an extended period of time. Explicit knowledge is formal knowledge that is easy to transmit between individuals and groups. It is frequently articulated in the form of mathematical formulas, rules, specifications, and so on. The two categories of knowledge are complementary. Tacit knowledge, while it remains closely held as personal know-how, is of limited value to the organization. Explicit knowledge does not appear spontaneously, but must be nurtured and cultivated from the seeds of tacit knowledge. Organizations need to become skilled at converting personal, tacit knowledge into explicit knowledge that can push innovation and new product development. Whereas Western organizations tend to concentrate on explicit knowledge, Japanese firms differentiate between tacit and explicit knowledge, and recognize that tacit knowledge is a source of competitive advantage. "

Figure 1 describes the several kinds of transformation processes between tacit and explicit knowledge. Nonaka & Takeuchi propose to understand organizational knowledge creation in a five-phase model which is closely related to these transformation processes.

The model (see Figure 2) comprises the following steps: (1) sharing of tacit knowledge, (2) creating concepts, (3) justifying concepts, (4) building an archetype, and (5) cross-leveling of knowledge. The several phases are explained as follows (Nonaka & Takeuchi, 1995):

`` ... The organizational knowledge-creation process starts with the sharing of tacit knowledge, which corresponds roughly to socialization, since the rich and untapped knowledge that resides in individuals must first be amplified within the organization. In the second phase, tacit knowledge shared by, for example, a self-organizing team is converted to explicit knowledge in the form of a new concept, a process similar to externalization. The created concept has to be justified in the third pase, in which the organization determines if the new concept is truly worthy of pursuit. Receiving the go-ahead, the concepts are converted in the fourth phase into an archetype, which can take the form of a prototype in the case of ``hard'' product development or an operating mechanism in the case of ``soft'' innovations, such as a new corporate value, a novel managerial system, or an innovative organizational structure. The last phase extends the knowledge created in, for example, a division to others in the division, across to other divisions, or even to outside constituents in what we term cross-leveling of knowledge. ...''

In the sequel, we will use one of Nonaka & Takeuchi's success stories in order to exemplify how IT may contribute to the successful application of the knowledge-creation model.

3 IT Support for Knowledge-Creation:
The Honda City re-invented

In 1978, the Honda top management inaugurated the creation of a new car concept with the slogan ``Let's gamble, let's develop a fundamentally different new car!'' and the only requirement to invent something "inexpensive but not cheap''. This stimulated a creative mood in the development team where vague information was freely exchanged and construc-tively elaborated.

Given our KARAT and EKI tool (to be presented in more detail in sections 4 and 5), the early phase of sharing tacit knowledge and, thus socialization among the employees would proceed in the following way. The team members communicate via a shared information space where arbitrary ideas or information represented in any kind of medium can be exchanged and discussed. These information sources may either float freely in the information space or may be connected by (typed or untyped) links among each other or to some categories they refer to (e.g. ``inexpensive, but not cheap'').

Since in a complex technical domain not each participant may be interested in all upcoming topics, the exchanged information items can be classified wrt. several organization schemes which allow to filter for manifold criteria. Such filters must be very flexibly configurable in order to cope with the complex and dynamically changing environment. E.g., it must also be possible to define personal organization schemes which allow, e.g. to weight discussion statements wrt. the relevance to my personal view of how close the topics are related to my own competence. In order to ease the use of the tool, it should be able to automatically classify the stored information items, at least wrt. some standard categories. In our Honda example, there is a ``Car_Specifications'' model with standard categories ``Performance'', ``Engine'', ``Driveline'', ``Dimensions``, ``Capacities'', and ``Chassis'' (see Figure 3).

Through this exchange of arbitrarily informal or vague information, unstated und usually intangible beliefs may become socially shared in the group. Besides the pure exchange of information, the tool also stimulates discussion between team members: comments, changes, or improvements can be attached to some information item, and the extended/changed item be put into the shared communication space, as well as directly e-mailed to its author. For face-to-face communication, an audio or video channel to an information item's author can also be opened. In this way, the socialization is promoted, and the represented ideas become evolved and improved step by step.

As discussion goes on, several new organization schemes may grow out of the shared information items, or the given organization scheme is changed or refined - just as in a brainstorming session using the METAPLAN method, when the participants suddenly become aware of some structure immanent to the topics on the note pads on the whiteboard. Evolving organization structures reflect the changing focus as well as progress of the discussion. In our Honda example, the new discussion tracks ``man maximum'' and ``machine minimum'' are introduced as subclasses of the ``fundamentally different'' category. Introduction of new organizing structures reflects social commitment within the group.

As the rough direction of the discussion is consolidated, the knowledge-creation process gradually passes over to the concept-creation phase. While the knowledge-sharing phase was characterized by ambiguity and redundancy, now processes become more goal-directed. Once focussed on certain aspects, an engineer's creativity may be stimulated by purposeful selection of all discussion contributions related to these aspects. This can be done by employing KARAT's retrieval facility which allows to extract all pieces of information attached to the selected model concepts. A construction engineer might be interested in all statements concerning the ``Chassis'' as well as the ``machine minimum'' classes. The system would present all information items which were generated within the ``machine minimum'' discussion and automatically classified to the ``Chassis'' standard category.

Since a prototype has to meet both the ``man maximum'' and the ``machine minimum'' ideas, now, a manager could be interested in possible hidden relationships between information items classified to either category. With the EKI tool, he lets search the information space by freely programmable ``information gathering agents'' (marker programs) which can find connections between information items which were not explicitly stated by the user. In the example, the markers start at the ``man maximum'' and ``machine minimum'' categories, respectively. From the ``man maximum'' category, one marker reaches an information unit ``big car'', from the ``machine minimum'' category, the other marker reaches a ``small car'' information unit. With the help of KARAT's text analysis component, both units have been automatically classified to the ``Dimensions'' category in the ``Car Specifications'' standard model. So, the marker passing mechanism detects this hidden relationship, and the EKI tool produces a summarizing statement "With respect to the ``Dimensions'' there seems to be an interesting relationship between the "small car`` and the "big car`` information units.'' Because of the apparent contradiction, the manager decides to inspect both information units in more detail. Using the KARAT information browser, the following explanations are presented:

Now, the manager could easily resolve the inconsistency by inventing the ``tall boy'' concept: a car which is large wrt. its height and ``small'' wrt. its length and width.

For the sake of space, we will not go into detail about tool support for the more conventional phases of concept justification, prototype building, and cross-leveling of knowledge. However, one can imagine how KARAT's selective examination of all information items relevant for some particular concepts or contexts eases the detection of inconsistencies, redundancies, or missing information. Similar to the scenario in concept creation, the EKI tool allows again to establish new views which have not been explicitly prepared in the KARAT models and contexts. This allows sophisticated procedures for checking complex consistency constraints wrt., e.g. business rules within the company. Since prototype building is often characterized by several trial-and-error cycles, using the same tool for supporting and documenting all phases allows continuous process improvement and requirements tracing in the case of changing conditions. In the cross-leveling phase, KARAT provides sophisticated storage and organization mechanisms for best-practice reports, whereas EKI provides flexible and powerful search across several conceptual spaces.

In summary, we sketched and illustrated how IT tools could comprehensively support Nonaka & Takeuchi's knowledge-creation process. Two main functionalities were identified:

In the subsequent sections, we will show how our KARAT and EKI tool, respectively, meet these requirements.

4 KARAT: Groupware Support for Collaborative Knowledge Elicitation, Sharing, and Organization

The KARAT tool and method were originally designed and implemented for preparing well-structured software requirements specifications from large amounts of unstructured, redundant, and inconsistent text documents as available from interviews with experts, minutes of meetings, e-mail discussions, etc. (Tschaitschian, Wenzel & John, 1997).

However, in (Tschaitschian, Abecker & Schmalhofer, 1997), we argued that the basic principles underlying this requirements engineering (RE) application could easily be applied to the broader context of enterprise knowledge management, and thus, the KARAT tool be re-interpreted as a knowledge management tool. To name the most important of these principles:

Recently, we added several groupware functionalities. Moreover, we are currently extending our information sources from pure text to arbitrary (multi-)media. This opens new application scenarios: the means for combining more or less informal knowledge sources and for flexible knowledge organization together with a powerful information sharing and communication concept allows a group to stepwise discuss and develop, formulate, and fix their shared knowledge.

Cooperating via Shared Information Sources

The KARAT tool offers to its users a shared information space where each participant in the knowledge management process is free to feed in arbitrary (multimedia) information items. Of course, tacit knowledge in general can not be communicated using a fixed representation medium and vocabulary. Consequently, the user should be allowed to forward text-based discussion contributions (e.g., figuratively described ideas on new car concepts) as well as graphics (e.g. the sketch of an innovative design detail), pictures (e.g., the scanned advertisements of a competitor's new car model), or even video tapes (e.g., driving studies with car prototypes or records of brainstorming meetings), and so on.

Technically, this multimediality is achieved by KARAT's HTML user interface (see Figure 3) which allows to handle arbitrary information sources using the appropriate Netscape plug-ins. Conceptually, the different knowledge media must be mapped onto some uniform meta-representation in order to exchange, compare, and discuss them. To this end, each information source is represented by a KARAT information unit (cp. Figure 4).

Representing Information Sources

As a first approximation, a KARAT information unit can be compared with the knowledge-item description frames proposed by (van Heijst, van der Spek & Kruizinga, 1996) for organizing Corporate Memories. There, knowledge items are described by their classification wrt. some fixed dimensions of knowledge categorization inspired by the KADS approach: tasks to be solved, role of the actor using the knowledge, application domain, etc. While such an approach gives a good idea of how a ``rough organization model'' (see above) could look like in an enterprise, it is nevertheless by far too rigid for tackling with the dynamics and unforeseeable effects in the early phases of Nonaka's model. While a fixed classification scheme may be useful for promoting reuse of stable and well-established procedures, i.e. knowledge conservation and documentation, for knowledge-creation, we need much more flexible, easily changeable, and multi-faceted descriptions. This goal is tackled in KARAT by two basic mechanisms:

A KARAT knowledge organization model is just a graphical notation of some coarse-grained, overall structure which can be used for categorizing information units.

Annotating Information Units

In the annotation part, an information unit is enriched by natural-language comments. In the RE application, this part contains, e.g. a decontextualized form, because single requirements extracted from larger text documents must be independently understandable. In the case of a multimedia information source, there can be, e.g. some short text description of the source's content, because it might be inconvenient to show a video tape to the user each time she inspects the information repository without roughly knowing whether it is of particular interest for her.

Relating Information Units

In the link part, hyperlinks can be drawn from each information unit to either arbitrary other units or arbitrary elements of knowledge organization models. We will refer to these two types of links as association links in the first and classification links in the latter case.

Association Links

Association links reflect the complex, network-like nature of knowledge which is always characterized by highly interwoven structures and close relationships between the respective information units. Association links can either be typed (representing some well-defined relation between information units) or untyped (giving only a rough hint for a ``has-to-do-with'' relationship). Examples for a typed association could be the ``part-of'' relationship between more complex requirements specifications and their several parts, or the ``follow-up'' relation between causally related discussion statements (e.g. the idea of a large passenger cabin is caused by the requirement to realize a ``man-maximum'' car). Untyped associations will be the predominant kind of links, because our application experiences show that the semantics of such a typed information link is hard to catch for a typical user. And, typed links are only useful in the case that there exist specialized utilization components which are able to exploit their semantics. So, we prefer to implement most of the wished structuring functionality by introducing additional (maybe individually tailorized) organization models which allow to produce specialized, small views on exactly that part of the knowledge network which is actually of interest.

Classification Links

Our organization models refine a certain view on the domain of discourse in a shallow way, by a few structuring classes (cp. Figure 3). In contrast to deep models as they are used, e.g. for model-based inferences, we need merely what (van Heijst, van der Spek & Kruizinga, 1996) call ``knowledge engineering at the macro level''.

Each information unit can be attached to arbitrary model components of any model. In a company, there will exist models of general importance as well as project-specific ones and models owned by single users or user groups. Information can be retrieved by selecting any combination of model components and specifying general information unit features (such as author, creation date etc.). So, arbitrary views and multi-criteria selections of the knowledge base can be produced. Since models can easily be extended or changed, or new models can be added, this organization scheme offers utmost flexibility.

The only remaining problem is that it is still some work to classify information wrt. the organization models. The first remark on this is that the user does not need to classify information units. If one does not use any model, the tool resembles a blackboard where everyone may attach some note pads in a completely unstructured way. However, when some user wants to structure the gathered notes, he is supported in a twofold manner: first, the process of classification and linking is very intuitive and comfortable to use through our carefully designed hypertext interface; second, the tool can provide suggestions (and can also make some classifications automatically) by the use of text analysis techniques coming from information retrieval. These suggestions are generated on the basis of keyword lists describing characteristic terms of the respective classes. The keyword lists as well as even suggestions for model classes can be generated with sophisticated algorithms from learning text categorization. In our Honda example, at least mapping onto some standard classes should easily be possible automatically.

Despite their ``formal'' appearance (in fact, our tool offers the graphical means for building many widely-used types of graphically denoted models, such as KADS inference structures, ER-Diagrams, semantic nets etc.), for the KARAT tool, most models are nothing else than flat sets of concepts. But this, together with the graphical presentation (the interpretation of which remains with the user) is already enough for enabling a very flexible, yet easy to use knowledge organization. Since this kind of ``shallow organization models'' is the most distinctive feature of the KARAT approach, we will elaborate a bit more on them.

Organization Models for Knowledge Structuring

KARAT employs the following types of organization models:

KARAT: Implementation

Figure 4 shows a screenshot of the KARAT user interface. The KARAT kernel system is implemented in Smalltalk. Up to now, two text analysis tools implemented in C and C++ (Dengel et al., 1994) are currently coupled to the KARAT kernel. MORPHIC-PLUS is a tool for the inflectional morphological analysis of the German language (Lutzy, 1995). The INFOCLAS2 system includes an automatic indexing component with different weighting functions as well as a word-based text classification component (Hoch, 1994). To guarantee user acceptance and user friendliness an integrative working environment has been realized through a uniform hypermedia user interface. The Hypertext Abstract Machine (HAM) is employed for hypertext management tasks, such as establishing and updating links between requirements and the different models. The Netscape Navigator Internetbrowser is employed as an interface for the HTML-represented information units and text retrieval results.

5 EKI: Tool Support for Creative Inferences

5.1 A Construction-Integration Approach to Knowledge Creation & Novel Frames of References

First, we describe a model of human learning which assumes that learning is a chaotic and globally unorganized process. The construction-integration model proposes that there are two phases of learning. In the first phase, the so-called construction phase, some new piece of information or idea (i.e. the learning material) is fully elaborated in combination with the already available knowledge. Many different kinds of associations which can be retrieved or constructed from the knowledge net (i.e. several knowledge bases) are formed. These knowledge construction processes may be performed under the guidance of a processing goal. They yield pluralistic views at various levels of abstraction.

In the second phase, a constraint satisfaction process determines the most dominant (e.g. the largest set) of knowledge units which are consistent with one another and as complete as possible with respect to some desired object domain. For example, it may be required that the resulting knowledge (the so-called situation-specific circumscription) presents a complete account of a possible design at some given level of abstraction.

As can be seen from Figure 5, the two phases correspond to two different skills, the knowledge construction skill and the knowledge integration skill. Both of these skills are shown as basic inference steps. For the construction skill, three input meta-classes are assumed. A learner's prior knowledge is organized in the form of a knowledge net (cf. Mannes & Kintsch, 1991). The learning materials are another input meta-class. In addition, the processing goal influences how the construction skill utilizes the information from the knowledge net and the learning materials.

The construction skill forms a multi-level representation with pluralistic views of the learning materials. One piece of text information may thus be represented from different perspectives and in different ways. These pluralistic views may furthermore contain opposite and contradictory information.

The integration skill then evaluates how the pluralistic views can be best fit together in a multi-level description, so that a coherent and consistent abstract structure is obtained. Thereby, the abnormalities (i.e. inconsistencies) that existed in the pluralistic views are eliminated or at least substantially reduced. The product of the integration skill is termed a situation oriented circumscription and presents the (relatively abstract) knowledge that is acquired from the learning materials in combination with the learners' prior knowledge.

The EKI-tool was developed within the framework of this construction-integration theory. This tool allows for the creation of new inference knowledge (Schmalhofer, Franken & Schwerdtner, 1997). EKI stands for the Evolution of Kreative Inferences (or Initiatives). With the EKI-tool an explicit inference statement is constructed by marker passing in a joint text- and knowledge base and a subsequent compilation process. In addition, this tool can be used to perform knowledge integration processes which can produce spatial, causal and other type of representations. EKI can be applied to construct reproductive as well as creative inferences in different kinds of knowledge bases which have become related to one another by so called crosslinks. With this system we can describe how the constraints that are provided from different information sources (like for instance a novel idea that is expressed by a brief text and the givens of an organizational structure, that is expressed by a knowledge map) may yield a deeper understanding wrt whether or not this novel idea can be implemented with the given organization. In the following section, we will describe the EKI-tool in some detail.

5.2 EKI: The Basic Approach

Conceptually, the EKI-System consists of two major processing components. One component constructs creative inferences from a knowledge net and inserts the created inferences into the knowledge net at proper locations. Because the knowledge net consists of a conglomerate of different knowledge segments and because the creative inference may yield additional inconsistencies, the resulting knowledge net will usually contain many inconsistencies and redundancies. It is therefore denoted as pluralistic views. The other component analyzes the resulting knowledge net for coherence and consistency and thereby selects an appropriately chosen subnet so that some global coherence and consistency criteria are met. The resulting knowledge net is called the situation-specific circumscription. In addition to these major components there is a component for creating the initial knowledge net (pluralistic views) from several well structured knowledge bases and some text and a component for presenting the pluralistic views to the user.

The starting point for the creative inferences is a number of separate knowledge bases, each of which is in itself well structured. For example the THINGS knowledge base consists of hierarchically structured object classes and respective subclass relationships. Each object is denoted by its name and several slots may be used for its description. The ACTIONS knowledge base similarly consists of hierarchically structured action classes. Because each of the different knowledge bases is a complete and correct representation of the respective segment of a domain, none of the knowledge bases itself affords creative inferences. An affordance for creative inferences arises by associating or linking nodes either between or within some of the knowledge bases. These links which may introduce alternative readings to the previously precisely defined representations are called crosslinks. Relating the different knowledge bases by such crosslinks yields the possibility for discovering new relationships which were not already implicitly contained in the original knowledge bases. After such crosslinks have been inserted the cohering knowledge base which was termed pluralistic views is obtained.

Creative inferences may now be constructed by focusing on two (or more) nodes, identifying relationships between these nodes and then compiling the discovered relationship into some explicit inference statement. This explicit statement is then called a creative inference and is subsequently inserted into the knowledge net. The creative inferences are in themselves additional crosslinks. The overall structure of the knowledge net is changed by these new links and some additional arbitrariness may thereby be introduced. In other words, the good structure of the original knowledge bases is dissolved to a certain degree and thereby affords the development of some new structure. The purpose of the integration process is to find such a new structure. This new structure is determined on the basis of some processing goal which identifies the nodes and types of links that are of central interest. By focusing on these nodes and links, a relatively coherent and consistent structure is then found, i.e. the situation-specific circumscription. The situation-specific circumscription, thus is a novel structure which emerges form the original knowledge bases, the cross-links (generated by the text) and the creative inferences which yield a shift of the focus in the knowledge net.

5.3 EKI-System Architecture

Figure 6 shows the architecture of the EKI system. On the left side there are the various information sources which are to be provided by the user. The knowledge base consisting of six segments (i.e. THINGS, IDEAS, ACTIONS, THOUGHTS, EVENTS, INSIGHTS) together with possible cross_links and the units of the TEXT_REPRESEN-TATION which are to be processed build the input to the system. The NEW_TEXT is entered into the system so that it corresponds unit by unit to the representational format of the TEXT_REPRESENTATION. There are two tools, which construct the initial Pluralistic_Views from the user's input (CreateNet) and present the result to the user (PresentToUser).

After this initial setup has been accomplished, creative inferences may be generated and a knowledge integration may be performed on the knowledge net (Pluralistic_Views). In order to explore creative inferences which may emerge from this knowledge net, the user may select two (or more) nodes of the net and provide some specification that serves as the basis for finding relationships between these nodes. After such relationships have been discovered by the system in terms of connecting paths between the nodes (that satisfy the user-supplied specification), another specification determines how some explicit statement is compiled from these paths and inserted into the net. All of these component processes are depicted in Figure 6 by the procedure denoted as Inferencing .

In order to generate the situation-specific circumscription by the knowledge integration processes, the user identifies one or several focal nodes of the net and provides a specification about which relationships are of interest and should therefore determine the situation-specific circumscription. The situation-specific circumscription is then calculated as that segment of the pluralistic views which satisfies the specified relationships. More specifically, the segment of the net consists of all the nodes that satisfy the given relationship with each of the focal nodes. These component processes are denoted in Figure 6 by the Integration procedure.

Inferencing

As can be seen from Figure 7, the inference construction is performed in two steps. Possible relationships between two (or more) nodes are identified by a marker passing procedure (Norvig, 1989). The resulting paths are then analyzed by a rule interpreter. The rules which are supplied to this interpreter specify under which conditions an inference may be constructed and how an explicit inference statement is to be compiled out from some given path.

Marker passing process: The purpose of the marker passing process is to find interesting connections and relationships between the particular nodes of the network. In order to find appropriate connections, the behavior of the marker can be programmed in terms of regular expressions. These regular expressions thus provide the specifications for which types of links can be traversed. In order to find paths between two nodes, a marker passing process is started from each of the two nodes. For every collision point of the two markers, a connecting path is then determined. Depending upon the particular specification that is used, the behavior of the markers may range from finding a very specific path over finding a concatenation of only transitive relations (e.g. like some transitive closure) to allowing any arbitrary connections between the nodes.

Inference compilation: After connection paths have been found, a rule interpreter is used for compiling explicit inference statements from a path and inserting these statements into the knowledge net. The rules for the rule interpreter consist of condition-action pairs. The condition part is again a regular expression which determines whether an explicit inference statement can be generated from a particular path by the action part of the rule. The action part then determines how the explicit inference statement is to be generated. Such an inference statement may consist of a single link that connects the two nodes from which the marker passing process was started. Alternatively, the inference may consist of a new node and two links, where the two links connect the new node to each of the two nodes from which the marker passing process was started.

Integration Processes

For performing the integration processes, the two or more nodes which are of focal interest must be identified by the user. Furthermore for each of these focal nodes, a marker program must be supplied which determines the relations (links) that are most significant and should therefore guide the integration process. As can be seen from Figure 8, the integration processes are performed in two steps. In the first step, the marker passing procedure is applied to flag all the nodes which are reachable from each given starting point by the respective marker program. In the second step, the integrated knowledge net (i.e. the situation-specific circumscription) is determined by some integration criterion. The integration criterion typically specifies that the situation-specific circumscription should be formed on the basis of those nodes which have been marked by all the different markers. There may be different criteria for specifying which links should belong to the integrated knowledge net. For instance, the situation-specific circumscription may include only those links which have been marked from all markers. Alternatively, it may include all the links that connect some nodes that have been marked. Obviously, there are several other possibilities as well.

5.4 EKI: Implementation

The EKI-system was designed in an object-oriented manner. The core components of the EKI-system are the knowledge bases, the marker passing process, and the inference compilation process. In the knowledge bases (knowledge net), one distinguishes nodes and links. All nodes are instances of object classes. There is one basic class, the BasicWorldClass, which has two subclasses, the StaticWorldClass which represents static entities and the DynamicWorldClass representing changes. By using Java, we programmed a user-interface, whose interaction style is very familiar from a large number of other applications in the WorldWideWeb. A demonstrator version of this system can be viewed over the net2). Because of the existing security restrictions of Java running under Netscape, the demonstrator cannot access any files. A regular version of the system, which allows full file access is also available. Since Java is an object-oriented and architecture neutral programming language, the system will be highly portable to almost any hardware platform.

The marker-passing procedure must be specified by the user, or an existing procedure must be selected. Figure 9 shows a screenshot of the EKI-system will the marker programs are specified that eventually lead to the inference of developing a short but tall cars. As can be seen from this figure, the marker programs are quite short, and as such are more readily entered by simply typing the program text than by the direct manipulation techniques used to perform other interactions with the EKI system. Important usability factors in the design of the marker-passing language include the regularity of the command structure, the number of commands a user must know and their mapping to the operations the user wants to perform. The complexity of this task is comparable with composing UNIX commands. The user can interactively run and re-run marker passing programs and the results can be traced back to terms in the program in a straightforward manner.

6 Discussion

Coming from the management point of view, Nonaka & Takeuchi recently presented their comprehensive model for understanding organizational knowledge creation. In order to cope with the uncertainty and dynamics of the upcoming knowledge society, the successful company must cultivate creativity and organizational flexibility. In this paper, we discussed how AI could adjust their techniques in order to meet these requirements. More specifically, we presented two tools:

The EKI tool supports the generation of creative inferences from a variety of knowledge sources. Because the separate entities of corporate knowledge need to be related, the generation of such creative inferences is indeed a necessary component in a learning organization. The tool is based on the construction-integration theory for understanding human learning processes. The KARAT tool, based on model-based knowledge acquisition, combines techniques from hypertext, text analysis, and groupware for sharing, evolving and gradually formalizing informal knowledge sources within a group of knowledge workers.

The primary aim of this paper was not to present the tools in technical detail (see (Tschaitschian, Wenzel & John, 1997) or (Schmalhofer, Franken & Schwerdtner, 1997) for this purpose), but rather to illustrate the scenario, to demonstrate their cooperation, and to discuss the main distinctive features relevant to the above question of how organizational learning could be supported. Some of the main topics to be further worked on:

First, handling of informal and tacit knowledge. We suggested that a combination of various knowledge media together with sophisticated means for group or face-to-face discussion may contribute to this goal.

Second, going beyond fixed frames of references and ``unfreezing'' or annealing established representation spaces. We showed how EKI may enable the user to do so. If learning is defined as a process which leads to an improvement in performance, and one which is intimately interconnected with memory (or knowledge), then the parallels between the individual and the corporate case emerge (Ayas and Foppen (1996) make a similar argument). It has also been argued that within organizations creativity occurs through interaction (Ayas and Foppen, 1996), and that learning occurs through discussion (de Geus, 1996). We propose that there are strong parallels between these social processes and the construction-integration processes of the cognitive model upon which the EKI tool is based. Learning in a changing world demands that contingent upon the external environment, new frames of reference must be dynamically constructed in a social effort in which all the practitioners of an organization are involved. These reconstructions are performed on the basis of innovation foci, which are analogous to a discourse focus in text comprehension (see van Elst & Schmalhofer, 1997). These innovation foci have to be identified by the knowledge managers of an organization. The EKI-tool supports the knowledge managers in this task.

Thirdly, both tools show that in early phases of knowledge creation, cooperative support systems and intelligent communication tools are more useful than fully-automatic problem-solvers. However, in order to get such tools really used in practice, AI has to care much more about psychological and ergonomic human-computer interaction aspects. This concerns not only the design of intuitive user interfaces, but also the conceptual design of the tool philosophy itself. Our customer contacts suggest that this is not only a question of how good a tool performs, but a critical success factor which is essential for a tool to be used at all. In KARAT, this is reflected by the thoroughly elaborated trade-off between formality (which allows powerful inferences and services) and informality (which makes the tool usable for a typical user). Issue-based information systems are a similar example where, for coordinating and documenting group decision processes in design, the relationships between information items are flexibly modeled by embedding them into an argumentation structure reflecting the discussion process. It may be useful to implement such a system with KARAT. In large, heterogeneous discussions which are still very unstructured an issue-base may provide a useful organization, similar to the various domain and task models in KARAT. In his EKAW-97 invited talk, Tom Gruber promoted a ``minimalistic'' approach to formality in corporate memories. He envisioned a self-organizing group knowledge creation process where useful organizing structures come ``on-the-fly'' as an emergent phenomenon; they are a by-product of people doing their knowledge work on-line from the way they name their public folders for starting and filing group discussions. Gruber's proposal shares many similarities with the approach of the KARATEKIT, which lies in between the two extremes of formalization. KARAT may mimic more or less formal behaviors depending on the defined models and typed links. A reasonable use of the tool is typically given with a handful, shallow organizational models and just two or three typed links, the semantics of which is easy to understand and can be exploited by specialized reasoning algorithms.

The last, and by no means less important points concern motivational and management considerations. In organizational knowledge creation, a creative mood in a group must be promoted such that individual creative acts are stimulated in relation to the body of shared group knowledge. Further, one needs to achieve and maintain the commitment of the employees to the socially constructed knowledge. Sensemaking and a changed explicit reward system are topics which are discussed for ensuring these points. If such goals are approached, a group of knowledge workers will definitely produce better results than the sum of the knowledge creations of the individual participants.

Acknowledgment. This work was funded by Deutsche Telekom AG and partially supported by the German Federal Ministry for Education and Research (bmb+f) under grant ITWM 9705 C4.

References

(Ayas & Foppen, 1996) Ayas, K. and Foppen, W. (1996) Reflections on design for learning.

(Boden, 1991) Boden, M. A. (1991) The creative mind: Myths and mechanisms. New York: Basic Books, Inc., Publishers.

(Choo, 1996) Choo, C.W. (1996) The Knowing Organization: How Organizations Use Information To Construct Meaning, Create Knowledge, and Make Decisions. Int. Journal of Information Management, 16(5), 329-340

(Clancey, 1991) Clancey, W. J. (1991). The Frame of reference problem in the design of intelligent machines. In K. VanLehn (Ed.), Architectures for intelligence. The Twenty-second Carnegie Mellon symposium on cognition (pp. 357-423). Hillsdale, NJ: Lawrence Erlbaum Associates.

(Dengel et al., 1994) Dengel, A., Bleisinger, R., Fein, F., Hoch, R., Hönes, F. & Malburg, M. (1994) OfficeMAID - A System for Office Mail Analysis, Interpretation and Delivery. Proc. of First International Workshop on Document Analysis Systems (DAS'94), 253-275, Kaiserslautern, Germany.

(Drucker, 1993) Drucker (1993) The Post-Capitalist Society. Butterworth-Heinemann Ltd.

(de Geus, 1996) de Geus, A. P. (1996) Strategy and learning.

(Gruber, 1993) Gruber, Th. R. (1993) Towards principles for the design of ontologies used for knowledge sharing. In N. Guarino and R. Poli (Eds) Formal Ontologies in Conceptual Analysis and Knowledge Representation. Kluwer Academic Publishers.

(Hoch, 1994) Hoch, R. (1994) Using IR Techniques for Text Classification in Document Analysis. Proc. of 17th International Conference on Research and Development in Information Retrieval (SIGIR'94), 31-40.

(Kidd, 1994) Kidd, A. (1994). The marks are on the Knowledge Worker. Proceedings of CHI-94: Human Factors in Computing Systems, Boston, Mass (24-28 April'94), 186-191. ACM Press: New York.

(Kintsch, 1992) Kintsch, W. (1992). A cognitive architecture for comprehension. In H. Pick, P. van den Broek, & D. Knill (Eds.), The study of cognition: Conceptual and methodological issues (pp. 143-164). Washington, DC: APA Press.

(Kintsch, 1994) Kintsch, W. (1994). Text comprehension, memory, and learning. American Psychologist, 49, 294-303.

(Kühn & Abecker, 1997) Kühn, O. & Abecker, A. (1997) Corporate Memories for Knowledge Management in Industrial Practice: Prospects and Challenges. J. Universal Computer Science. 3(8), 929-954. Springer Science Online.

(Lutzy, 1995) Lutzy, O. (1995) Morphic-Plus: Ein morphologisches Analyseprogramm für die deutsche Flexionsmorphologie und Komposita-Analyse. DFKI Document D-95-07.

(Mannes & Kintsch, 1991) Mannes, S., & Kintsch, W. (1991) Routine computing tasks: Planning as understanding. Cognitive Science, 15, 305-342.

(Nonaka & Takeuchi, 1995) Nonaka, I., & Takeuchi, H. (1995) The knowledge-Creating Company. Oxford: University Press.

(Norvig, 1989) Norvig, P. (1989). Marker passing as a weak method for text inferencing. Cognitive Science, 13, 569-620.

(Schmalhofer, 1997) Schmalhofer, F. (1997) Constructive Knowledge Acquisition: A Computational Model and Experimental Evaluation. Hillsdale: Lawrence Erlbaum Associates.

(Tschaitschian, Abecker & Schmalhofer, 1997) Tschaitschian, B., Abecker, A. & Schmalhofer, F. (1997) Information Tuning With KARAT: Capitalizing on Existing Documents. In E. Plaza & R. Benjamins (eds.) Knowledge Acquisition, Modeling and Management, 10th European Workshop, EKAW`97, Berlin: Springer, LNAI 1319, 269-284.

(Schmalhofer & Tschaitschian, 1995) Schmalhofer, F.& Tschaitschian, B. (1995) Cooperative knowledge evolution for complex domains. In G. Tecuci & Y. Kodratoff (eds.) Machine Learning and Knowledge Acquisition: Integrated Approaches, London: Academic Press, 145-166.

(Schmalhofer, Franken & Schwerdtner, 1997) Schmalhofer, F., Franken, L., & Schwerdtner, J. (1997) A computer tool for constructing and integrating inferences into text representations. Behavior Research Methods, Instruments, & Computers., 29 (2), 204-209.

(Schmidt, 1995) Schmidt, G. (1995) Modellbasierte, interaktive Wissensakquisition und Dokumentation von Domänenwissen (MIKADO), DISKI Vol. 90. Sankt Augustin, Germany: Infix.

(Shum, 1997) Shum, Simon Buckingham (1997). Negotiating Multidisciplinary Integration: From Collaborative Argumentation to Organisational Memory. J. Universal Computer Science. 3(8), 929-954. Springer Science Online.

(Tschaitschian, Wenzel & John, 1997) B. Tschaitschian, C. Wenzel, and I. John. Tuning the quality of informal software requirements with KARAT .. In E. Dubois, L. Opdahl &K. Pohl (eds.). REFSQ'97: Third Int. Workshop on Requirements Engineering: Foundation for Software Quality, Held at CAiSE*97, Barcelona.

(van Heijst, van der Spek & Kruizinga, 1996) van Heijst, G., van der Spek, R. & Kruizinga, E. (1996) Organizing Corporate Memories. In Tenth Knowledge Acquisition for Knowledge-Based Systems Workshop KAW'96. Banff, Canada..

(Wielinga, Schreiber & Breuker, 1992) Wielinga, B. J., Schreiber, A. T., & Breuker, J.A. (1992) KADS: A modeling approach to knowledge engineering. Knowledge Acquisition, 4 (1), 5-53.


Last Modified: 06:26pm MET, November 03, 1997