The idea that knowledge can be used for finding information through the network has been developed in several directions. However, most proposals keep the knowledge about the information separate from the access methods. This has the advantage of making it possible the development of systems that can gather information from different sources, but it requires a mapping from the representation structures into methods to access information on the network which is accomplished by ad hoc procedures. On the other hand, the information that we access through the browsers is organized into structures that are coded in terms of the content of each page and of its relationship to other pages. Such structures typically guide humans in the search for a specific information. Our idea is to use this kind of structural information to build knowledge based tools for the access to the network. More specifically, we are working at a system which acquires information on a subject matter by extracting from the pages an explicit representation of the underlying structure, and use it to answer user's requests by directly pointing to the page containing the desired answer.

There are three building blocks of our proposal: a mechanism for representing Web pages in terms of description logics, a method for learning the structure of the pages available in the network on a specific topic, a technique for generating a plan, as a sequence of links, that leads to a page satisfying a user request.

Description logics [4] have been developed in the last years with the aim of formalizing terminological systems that have been developed after KL-ONE. Such systems are based on the idea of representing knowledge by building hierarchical structures, sometimes called terminologies, where the descriptions of interest are classified according to their properties. The basic elements of the representation are called concepts and roles, which denote classes and binary relations, respectively. This general scheme is specialized in our framework according to the following ontology:

• the structure of a page is regarded as a concept description and a specific Web page as an instance of a concept;
• each page has static properties, which are represented as roles and refer to any kind of information that can be relevant to both the page structure and its content;
• each page has dynamic properties, which are also represented as roles, and are represented as actions, in the sense that they allow jumping to other pages.

Regarding the learning method, the extraction of information from the sources relies both on the ability to represent the syntactic structure of the documents (description logics can be effectively used to characterize SGML documents [1]) and on the processing textual information in order to identify names for the links and other static properties such as, for example, page title. The information acquired by the system is returned as a set of training examples in a description logic. To support the construction of a knowledge base, we use the method for inductive learning proposed in [2]. In addition, we are considering the extension of such method to exploit the use of background knowledge during the learning process. The result of the knowledge acquisition stage is a knowledge base containing the concept descriptions that represent the typical structure of the Web pages on the subject matter, together with information about the initial addresses to navigate within such structures.

Once the knowledge base is constructed, our system can accept requests from the user, concerning the subjects previously examined, and is expected to return the page(s) satisfying the requests. This is achieved by a two step process. The first step consists of finding one or more plans (sequences of links) that lead to a page satisfying the goal associated with the user request. The stecond step consists in the execution of the plan, which may require the discovery of failures and requests for replanning. The planning component of the system relies on a special use of the procedural rules typically used in description logic systems, that has been obtained by exploiting the correspondence between description and dynamic logics. The setting has been developed for planning the actions of a mobile robot [3].

Therefore, our system, whose implementation is in progress, is meant to act as an intelligent browser, which allows the user to reach the requested information in a single step. Learning and planning are essential underlying techniques for accomplishing such a task, and description logics are the bases for both the formalization and the implementation of several system components. We believe that the main feature of our proposal is in the intrepretation of the network structures available on the Web, directly in terms of the representation structures used by the system in order to provide reasoning facilities.

References

[1] D. Calvanese, G. De Giacomo, M. Lenzerini. Rethinking SGML document type definitions

[2] W. Cohen, H. Hirsch, Learning the Classic Description Logics: Theoretical and Experimental Results. Proc. of KR-94

[3] G. De Giacomo, L.Iocchi, D. Nardi, R. Rosati. Classing Planning for Mobile Robots

[4] W.A. Woods. Understanding Subsumption and Taxonomy. A Framework for Progress. In Principles of Semantic Networks. John F. Sowa (Ed.)