Tasks as context for intelligent agents

The context of a statement or assertion can be critical to a useful understanding of its meaning. In huge knowledge bases such as the CYC project, capturing context explicitly encounters significant problems. Research at the Intelligent Agent Lab at The University of Melbourne has been investigating lightweight approaches to incorporating context in the construction of knowledge-based information agents. The prototypical example of a successful agent built in the lab is SportsFinder. SportsFinder extracts sporting match results from a large variety of web sites without any formal knowledge representation. This is a task for which understanding the context of knowledge on the web pages is critical, as the individual elements that make up a page of sports results are very ambiguous without consideration of the surrounding information. This paper advocates incorporating context via a task specification, without resort to general purpose knowledge techniques.