Andreas Abecker, Ansgar Bernardi, and Michael Sintek
German Research Center for Artificial Intelligence (DFKI GmbH)
Knowledge Management Group
ABSTRACT: Proactive knowledge delivery, context-sensitive
information retrieval, dealing with heterogeneity in manifold aspects,
and self-adaptiveness are crucial requirements for innovative enterprise
information infrastructures supporting comprehensive enterprise knowledge
management (KM). We illustrate the services offered by the KnowMore system,
a prototypical realization of at least the first three of these principles.
We sketch the system architecture and some implementation issues and present
directions for the future work in this area.
1. ORGANIZATIONAL MEMORY INFORMATION SYSTEMS
Today?s companies exist in a world where markets are continuously shifting, technologies proliferating, competitors multiplying, and products become obsolete overnight. In such highly dynamic environments managers increasingly recognize knowledge as one of the most decisive business factors to deal with the continuous change. Knowledge management (KM) is a strong trend in the management sciences that must be appropriately supported by business information systems and enterprise information infrastructures (see, e.g., (WIIG, 1993), or (NONAKA and TAKEUCHI, 1995)). However, besides this general awareness of the issue?s importance, it is not clear at all what specific contributions are expected from the information technology (IT) and information systems (IS) areas, or whether a comprehensive picture of IT for knowledge management exists. One reason is that the KM hype is not fueled by a one single cause but is merely driven by a number of current business phenomena, like, e.g., lean management, which cut off the middle management from many organizational structures. For decades, just this level of employees had been the most important stakeholder of know-how and experience about projects, customers, and products, that used to perform such important tasks as information gathering, filtering, interpretation, and routing in order to prepare the decisions of the upper management. Another example is that the developed countries of the North shift more and more from producing goods to providing knowledge-intensive services. This brings into the center of interest knowledge-intensive work processes and wicked-problem solving (cp. (DAVENPORT et al., 1996), (CONKLIN and WEIL, 1997)).
Manifold such business phenomena produce a whole bunch of KM related aims comprising:
2. THE KNOWMORE KNOWLEDGE SERVICES AT A GLANCE
As an example consider the business process for purchasing goods in our research institute. Figure 1 gives an impression of the formal workflow to be performed represented with the ADONIS commercial business process modeling tool (cp. (KARAGIANNIS et al., 1996), (BOC GmbH, 1998)): A purchasing process starts with an employee filling out a demand specification form and mainly consists of a fairly deterministic sequence of more or less simple administrative steps like checking the budget, writing the order, or assigning an inventory number.
However, among these "primitive" administrative things, there are a few working steps requiring expert knowledge and purchasing experience. Some of them are marked in the picture by a dark surrounding circle. We will focus here on the preparation of a detailed specification of the goods to be purchased (Which model from which producer delivered by which supplier?) based on the more or less concrete demand description of the employee who initiated the purchasing process (I need some high-end PC with a good graphics card!). If an unexperienced employee should accomplish such a detailed demand specification, questions like the ones shown in Figure 1 could arise the answering of which would be a helpful service of an OM system.
In order to provide such a service, the KnowMore approach proceeds with the following steps:
Figure 1. A typical purchasing workflow.
Figure 2 shows a screenshot of our experimental system prototype. On the left, in the background, we see an editor window of the workflow application used to create detailed demand specifications. The input mask accepts up to three items to be purchased and already contains the initial specification given by the end user (she needs one graphics card, in German Grafikkarte). Now, it has to be decided which concrete card to buy (the product slot in the input mask) from which supplier (the supplier slot). The KnowMore system supports this decision in the following way:
When the workflow engine starts this activity, the system takes the information needs associated to the activity and finds out whether some element of the OM can already compute a decision suggestion (i.e., whether there is some expert system or decision support functionality available which can readily be evaluated). This suggested decision value is inserted in the user input mask offering a proposed solution (in the example, the suggestion is to buy a Matrox Mystique card).
Moreover, the system determines about which decision variables the system can offer some information, and inserts information buttons ("I") at the appropriate places in the input mask. If the user wants some supporting information on one of these decisions, pressing the "I"-Button starts a query to the OM system. This query may retrieve several classes of information (e.g., highly recommended company-internal business rules for purchasing in general or for specific product classes, technical information about possible buying alternatives, or pointers to knowledgeable colleagues who are known to be competent for such kinds of decisions because of their entry in the personal skill database, their training records or position in the company, or because they recently did a similar purchase) with different relevance in the concrete situation. These information sources are ordered according to their relevance computed by the retrieval function as well as to a predefined order based upon their information type, and are offered to the user as hyperlinks in the KnowMore information browser (Figure 2, right hand side).
Figure 2. The KnowMore system offers context-sensitive information supply.
The user can either accept or overwrite a suggested solution, and-in the case of a more detailed partial specification or a changed situation-again ask for KnowMore support via an "I" button. The system recognizes the change of the state of affairs and reevaluates the query against the archive system. Figure 3 shows the effect of taking into account the change in the process state: in the right part of the picture, the user has selected the Matrox Mystique card which considerably narrows down the search in the OM. Now, all documents which have no direct relationship to this specific card are eliminated. What remains are only compulsory purchasing business rules and specific information about the Matrox Mystique product. If we would ask for information support concerning potential suppliers, the system would yield only information about suppliers which are known to sell the Matrox Mystique while in the previous process state (left hand side of Figure 4) all suppliers would be described selling graphics cards in general. For realizing the functionality described, the KnowMore system makes the following technical provisions:
Figure 3. Changes in the process status result in refined information support.
3. REALIZATION OF THE KNOWMORE SYSTEM
Since a most innovative feature of our architecture is the workflow-triggered activation of information supply which exploits workflow context for precise search, we elaborate a bit on the question how to achieve the workflow-retrieval coupling, i.e., how to model KIT support specifications.
A KIT forms a unit of sense, but further details might be given by specifying several (partial) information needs. Each information need will result in some information which supports a particular aspect of the complete KIT. If there are logical or time dependencies among the output of several information needs known at process definition time, they can be represented in the KIT description. To this end, the KnowMore system provides preconditions in the various information needs, and processing rules for their results. Both influence the way the information needs are interpreted and fulfilled during process execution. If all relationships and interdependencies between information needs were known, we could probably represent the KIT as an ordinary workflow. But this is per definitionem not possible for a KIT. Hence, the imposed structure is only partial; it will be used for more effective information search and presentation, and not for guiding and controlling KIT processing.
As mentioned above, a KIT is a special case of an ordinary workflow activity extended by a support specification (containing information needs and processing rules) which may refer to the global and local process context (the lower part of the task box in Figure 4). The support specification specifies:
Figure 4. Information needs refer to the current global and local process status.
Figure 5. The three-layered OM architecture.
Concerning KIT processing, it must be stated that the central instance to work on the KIT is still the human workflow participant. He is responsible for solving the problem at hand. The worklist handler simply presents to the workflow participant an editor window with the KIT name and the input and output variables. The human user solves the task at hand by filling the output variables (at least in our current demonstrator system). In parallel, the KIT representation is passed to the info / knowledge agent which evaluates the information needs and instantiates the parameters. It then presents the various information needs as support offers (e.g., "I" buttons) to the user, using the name and the comment of the information needs (cf. Figure 6). The user selects interesting offers. Then, the info agent determines the relevant information sources, creates suitable queries from the information need, and performs the information retrieval. The result is presented as supporting information to the user.
Any change in the various variables which the user has to fill must result in a re-evaluation of the information needs which depend on these variables. This again shall be realized as a suggestion to the user: The previous results are marked as possibly outdated, but the activation of a new information retrieval is left to the user. As soon as the user completes the task and the filling of output variables, a message is passed to the worklist handler (as already indicated in the generic model by the workflow management coalition). Automatically the knowledge agent receives a close signal for this particular KIT, closes the display windows under its responsibility, and exits.
Figure 7 gives an overview of our current implementation. The KnowMore prototype is implemented in JAVA which allows it to be used on all JAVA-enabled platforms. The KnowMore server holds all relevant data, i.e., the business process model enriched by KIT variables and support specifications as well as the OM archive together with the respective knowledge descriptions and the underlying ontologies. Business process models can be designed using the ADONIS commercial BPM tool (the KnowMore specific extensions are modeled as comments in the activity descriptions), and are later parsed into the KnowMore representation formalism. The core of all so-represented knowledge is an object-centered knowledge representation formalism. The basic language constructs of this formalism are mapped in turn onto conventional relational databases which are coupled with JAVA via JDBC.
The KnowMore server hosts both the workflow engine and the knowledge-based retrieval machinery. Workflow enactment involves two parts: the server, implemented as a JAVA application, and client worklist handlers, implemented as JAVA applets which connect to the server via standard TCP/IP sockets. The architecture and communication protocols are designed in compliance with the Workflow Management Coalition standards such that later on, when the scenario is stable and proved, it should be possible to switch from our homemade KnowMore workflow engine to a commercial one.
Figure 6. KIT processing by worklist handler and info / knowledge agent.
The KnowMore system is already running, but several points
still have to be finally fixed; in particular the several declarative formalisms
for knowledge description, support specifications, and search heuristics
exploiting the different ontologies. Even if the technical points are clarified,
practical experiences must be gathered concerning an appropriate usage
methodology (when and how to acquire the necessary KIT descriptions and
support specifications). We think about a methodology for business-process
oriented knowledge management aiming at an intertwined modeling of processes
and KITs supported by an extended BPM tool like ADONIS. In practice, such
a comprehensive methodology plus additional tool support (e.g., providing
a library of info agents) seems indispensable, since the several modeling
activities required for implementing a KnowMore like system in a company
will consume considerable resources.
Figure 7. The KnowMore implementation is a Web-enabled client-server architecture.
One of those fundamental modeling questions concerns the ontologies used for information modeling. There seem to be some knowledge structures applicable in a wider range of applications (e.g., (O?LEARY, 1998) identified quite similar repository organization principles used in the most large international consulting companies). Nevertheless, building ontologies is costly, and it is not clear, in the general case, how stable they are, how detailed they should be, and whether they will be accepted by employees for indexing their knowledge. Here is much room for empirical studies grounding on a whole body of work done in the areas of knowledge acquisition and knowledge-based systems. It should also be noted that in a real company the knowledge acquisition process for building those ontologies has not necessarily to start completely from scratch. In every case you have already some company handbook, some company Intranet, some documentation system or company library, there exist already organization structures which can be used as an input for building the respective ontologies. In our KnowMore project we also investigated statistical text analysis techniques for finding co-occurences of words in given document corpora as an input for ontology construction and extension. Furthermore, for finding formal knowledge descriptions to index informal text documents, we experiment with information extraction and learning text categorization approaches from shallow text analysis.
Regarding the three OM requirements given in the introduction
section, this paper focused on the active, context-sensitive knowledge
supply realized by workflow-triggered and embedded information retrieval.
The integration issues are to some extent addressed in our knowledge description
level which was not discussed in great detail here. Self-organization and
self-adaptivenes are still the most challenging topics in OM research.
However, the just mentioned research in automatic text categorization and
ontology construction support go into this direction.
The work described in this paper has been done in the project KnowMore funded by the German Federal Ministry for Education and Research (bmb+f).
ABECKER, A., BERNARDI, A., HINKELMANN, K., KÜHN, O. and SINTEK, M. (1998). Towards a Technology for Organizational Memories. IEEE Intelligent Systems, 13(3).
BOC GmbH (1998). The Business Process Management Toolkit ADONIS. URL http://www.cso.net/boc/english/index.htm
CONKLIN, E.J. and WEIL, W. (1997). Wicked Problems: Naming the Pain in Organizations. White Paper of Group Decision Support Systems, Inc. URL http://www.gdss.com/wicked.htm
DAVENPORT, T., JARVENPAA, S.L. and BEERS, M.C. (1996). Improving Knowledge Work Processes. Sloan Management Review. Reprint Series 37(4).
KARAGIANNIS, D., JUNGINGER, S. and STROBL, R. (1996). Introduction to Business Process Management Systems Concepts. In Business Process Management (SCHOLZ-REUTER, B. and STICKEL, E. Ed.). LNCS, Springer-Verlag.
KÜHN, O. and ABECKER, A. (1997). Corporate Memories for Knowledge Management in Industrial Practice: Prospects and Challenges. Journal of Universal Computer Science, 3(8), Special Issue on Information Technology for Knowledge Management, Springer Science Online.
LIAO, M., HINKELMANN, K., ABECKER, A., and SINTEK, M. (1999). A Competence Knowledge Base System for the Organizational Memory. In XPS-99: Knowledge Based Systems - Survey and Future Directions, 5th Biannual German Conference on Knowledge Based Systems, Würzburg, Germany (F. PUPPE Ed.), Lecture Notes in Artificial Intelligence 1570, Springer-Verlag.
NONAKA, I. and TAKEUCHI, H. (1995). The Knowledge-Creating Company. Oxford University Press.
O?LEARY, D. (1998). Using AI in Knowledge Management: Knowledge Bases and Ontologies. IEEE Intelligent Systems, 13(3).
STEIN, E.W.and ZWASS, V. (1995). Actualizing Organizational Memory With Information Technology. Information Systems Research, 6(2).
WIIG, K.M. (1993). Knowledge Management: Foundations. Schema Press, Arlington, Texas.