The area of Knowledge-Based Information (KBIR) is an active one, with IR researchers trying to apply various AI schemes to improve IR systems, and with AI researchers trying to extend their systems and apply their methods in the domain of IR. This latter tact involves not only traditional IR, but also fact extraction, where large document collections are analyzed (either in bulk, or in connection with message or document routing tasks) for particular types of information, so that a database or fact base can be loaded with the extracted particulars.
From the IR perspective, KBIR is still unproven. In limited domains, for small collections, it seems to have particular promise, especially when specialized knowledge bases can be constructed to help with particular tasks or problems (e.g., accurately selecting search terms). KBIR also seems to help build systems that are more usable or efficient, where the knowledge is reflected in the system behavior. In part this is the tact taken by Belkin et al. in developing Distributed Expert-Based Information Systems (DEBIS) [1]. However, it is still unclear if and how KBIR can be applied to the important problem of ``understanding'' large numbers of documents in a heterogeneous collection, with truly robust analysis, and with significant improvements in effectiveness when compared to the best statistically-based systems.