ASPASIA Knowledge Systems - Knowledge Broker: Knowledge Generation and Handling

The Know-Broker tool exploits context related structures by evaluating user defined rules about knowledge object relations resp. attributes and corresponding concepts. As a base unit the user defined rule is represented as a statement for defining essentials (knowledge object features and characteristics) and / or constrains for a special chosen knowledge domain.

As semantic middleware component (like a semantic based information routing component) the Know-Broker integrates heterogeneous data and information sources and acts as the interface between the user and the unstructured knowledge pools.

Furthermore the information and knowledge objects are represented in semantic based ontologies (DAML), which ensures characteristics for machine readability and agent accessibility.

Due to the fact that the Know-Broker enables "intelligent" (inferencing procedures) and adaptive ("learning from the past") processing of knowledge objects, it generates "new" knowledge, based on associative and similarity oriented relations and inference based anotations.

The suchlike generated OTVs representing the diverse knowledge/information objects are processed in further different procedures according to there progress stage and properties:

Base condition for this different procedures is the determination of object similarity. For calculation of overall similarity matrices a modified combination of the DICE and TFIDF scheme is used.

Based on this similarity calculation the knowledge objects are investigated and used either within the clusterer stage or the trainer stage to generate new or predefined classes. In the following this main procedures are shortly illustrated:

The three-stage clusterer is used to generate specified object classes out of the collection of acquired (unknown/new) knowledge objects, which are repre-sented by object term vectors (OTVs). A three stage bottom up clustering is processed by using stepwise the MinMax-algorithm and proc-essing the similarity matrix of the objects given in the collection. For each generated class a centroid term vector (CTV) is calculated out of the OTVs referenced to the specified class. Accordingly the CTVs represent the nodes of a three-layered object/subject tree. The overall clustering stage yields in coupled subject classification trees, whereby their relations are expressed by relationship functions.

In the case of processing objects known to the administrator a trainer is realized, which generates and determines subject classes out of categories of predefined objects. This classes are furthermore referenced to the subject classification tree, whereby each node is again represented by a CTV.

The intrinsic classification of new - previously unknown - objects consists in the first stage of an estimation of a relationship coefficient between the OTV of the new object and the existing normalized term vector of the centroids. The resulting relationship vector – consisting of the calculated similarity factors – is a measure for the degree of relationship between a group of new objects and the centroids. A further three layered MinMax clustering of the relationship vector leads finally to three (at least two) groups of objects belonging to determined centroids representing the classes within the classification tree.