A mediator in agent-based architectures is a flexible component that provides a currently missing link between a lot of interesting data (which resides on the Web) and the applications (which have no direct access to the Web data), using wrappers and facilitators.
Efficient data extraction, storage, retrieval, transformation or modification, and reasoning are major issues when setting up an implementation strategy for a mediator that shall enable constructive reasoning re-using Web data. This paper has proposed a new representation framework for designing a mediator. It is based on labelled DAGs augmented by constraints as combined model for representing syntactic patterns, structure and shared semantically meta-descriptions of Web resources. Labelled DAGs provide a cognitive and natural approach to (1) represent static structure of XML Web knowledge bases and DTDs, (2) to model knowledge sharing, and (3) to automate query-processing, transformations and constructions. Arithmetic, set-theoretic or boolean constraints are essential for modelling system components with respect to synthesis and design tasks. Neither current XML standards nor graph-based models for Web data support the use and exchange of constraints. In [Eus97], canonical term representatives for isomorphic DAGs and a graph term algebra for implementing various operations on graphs were devised. These results allow storing structured graph representations of Web resources as terms embedded into logical knowledge specifications. Constraints to be applied to the structure or content of Web resources can be described as constraints attached to the graph term variables. Certain operations for retrieval, modification and reasoning that might otherwise have been performed by either combinatorial search or explicit specialised algorithms can be performed in a more efficient way by mechanisms that depend on the structure of the representation and using the built-in algorithms for constraint-solving, unification or subsumption of a constraint-logic programming language. Wrappers which translate XML data and meta-data and XML-QL queries into logical structures can be easily built.
There has been some interest recently in moving data from the WWW into knowledge that can be re-used by intelligent applications. Roughly related projects in this area use an ontology approach for specifying the domain of discourse as in HERMES [A95], and declarative languages for specifying mediators and performing reasoning for the automatisation of mediation tasks as in KOMET [CJKS97]. ONTOBROKER [ES00] uses ontologies that are formalised in Frame Logic to derive DTDs, whereas the presented concept of lifting rules provide a flexible means for the linking of independent ontologies and DTDs. [Kel95] presents an agent-based architecture for smart catalogues and brokering that supports cross-search of multiple heterogeneous SQL-based product data bases. Ontologies are used to define common terms. The KRAFT (Knowledge Re-use and Fusion Transformation) architecture [PHG99] focuses on the utilisation and re-use of constraint knowledge held in distributed sources, by transforming it for use with respect to various ontologies using a non-executable interchange format.
Future work will focus on the loose integration of concrete common ontologies and Web resources as knowledge bases into the SEAMLESS synthesis and learning system [Eus00]. The graph-based methodology itself can be extended to other semi-structured information bases.