A big question, that has not yet received much attention, concerns the extent to which retrieval effectiveness is limited by the type of document description used. The use of keywords to describe documents has affected the way in which the design of an automatic classification system has been approached. It is possible that in the future, documents will be represented inside a computer entirely differently. Will grouping of documents still be of interest? I think that it will.

Document classification is a special case of a more general process which would also attempt to exploit relationships between documents. It so happens that dissimilarity coefficients have been used to express a distance-like relationship. Quantifying the relationship in this way has in part been dictated by the nature of the language in which the documents are described. However, were it the case that documents were represented not by keywords but in some other way, perhaps in a more complex language, then relationships between documents would probably best be measured differently as well. Consequently, the structure to represent the relationships might not be a simple hierarchy, except perhaps as a special case. In other words, one should approach document clustering as a process of finding structure in the data which can be exploited to make retrieval both effective and efficient.

An argument parallel to the one in the last paragraph could be given for automatic keyword classification, which in the more general context might be called automatic 'content unit' classification. The methods of handling keywords, which are being and have already been developed, will also address themselves to the automatic construction of classes of 'content units' to be exploited during retrieval. Keyword classification will then remain as a special case.

H. A. Simon in his book The Sciences of the Artificial defined an interesting structure closely related to a classificatory system, namely, that of a nearly decomposable system. Such a system is one consisting of subsystems for which the interactions among subsystems is of a different order of magnitude from that of the interactions within subsystems. The analogy with a classification is obvious if one looks upon classes as subsystems. Simon conceived of nearly decomposable systems as ways of describing dynamic systems. The relevant properties are (a) in a nearly decomposable system, theshort-run behaviour of each of the component subsystems isapproximately independent of the short-run behaviour of the othercomponents; (b) in the long run, the