An Experiment in Integrating Internet Information Sources

The number of online information sources is growing rapidly. Though much of this information is unstructured (e.g., text, images) the number of structured information sources (e.g., databases) is also increasing. The advantage of structured information sources is that we can actually query their contents and use them to answer queries, perhaps in combination with other structured sources, rather than just viewing the whole information source. Given the variability in the information sources available, it is impractical for a user to interact with each source using its specific terminology. As a consequence, several systems (such as Nomenclator [Ordille and Miller, 1993b; Ordille, 1994], The Information Manifold [Kirk et al., 1995], TSIMMIS [Chawathe et al., 1994], SIMS [Arens et al., 1994]) have been developed based on the notion of a mediator [Wiederhold, 1992]. The key idea behind a mediator is that a user interacts with one uniform model of the domain, and the mediator translates between queries posed in the domain model, and the ontologies of the specific information sources. In order to do so, the mediator needs descriptions of the specific information sources that relate their contents to the domain model seen by the user. A major open issue which is crucial to the success of applying the notion of a mediator to large numbers of networked information sources is the ability to obtain accurate descriptions of the sources. The difficulty lies in problems such as sources using different attribute names to refer to the same intended attribute, or using the same attribute name to refer to different intended attributes. Further complications arise when the format of the attributes differs, even if they mean the same thing. An important characteristic of many networked information sources is that they fall into classes that model very similar domains. For example, a large class of information sources describe data about people in an organization, such as their addresses, phones, and email addresses. Similarly, many information sources contain bibliographic citations, or product and price listings. This paper describes an approach to the problem of obtaining descriptions of information sources that is well suited for networked information sources of this type, and the first large scale experiment that we are currently conducting, aimed at obtaining such descriptions in a real world environment. Our approach is based on the assumption that obtaining descriptions of information sources is a task that will inherently involve the managers of the information sources themselves, and the key to the solution of the problem is to find ways in which they can provide the most accurate description of their sources with minimal effort.