The time is ripe for another attempt at using natural language to represent documents inside a computer. There is reason for optimism now that a lot more is known about the syntax and semantics of language. We have new sources of ideas in the advances which have been made in other disciplines. In artificial intelligence, work has been directed towards programming a computer to understand natural language. Mechanical procedures for processing (and understanding) natural language are being devised. Similarly, in psycho-linguistics the mechanism by which the human brain understands language is being investigated. Admittedly the way in which developments in these fields can be applied to IR is not immediately obvious, but clearly they are relevant and therefore deserve consideration.

It has never been assumed that a retrieval system should attempt to 'understand' the content of a document. Most IR systems at the moment merely aim at a bibliographic search. Documents are deemed to be relevant on the basis of a superficial description. I do not suggest that it is going to be a simple matter to program a computer to understand documents. What is suggested is that some attempt should be made to construct something like a naïve model, using more than just keywords, of the content of each document in the system. The more sophisticated question-answering systems do something very similar. They have a model of their universe of discourse and can answer questions about it, and can incorporate new facts and rules as they become available.

Such an approach would make 'feedback' a major tool. Feedback, as used currently, is based on the assumption that a user will be able to establish the relevance of a document on the basis of data, like its title, its abstract, and/or the list of terms by which it has been indexed. This works to an extent but is inadequate. If the content of the document were understood by the machine, its relevance could easily be discovered by the user. When he retrieved a document, he could ask some simple questions about it and thus establish its relevance and importance with confidence.

Future developments

Much of the work in IR has suffered from the difficulty of comparing retrieval results. Experiments have been done with a large variety of document collections, and rarely has the same document collection been used in quite the same form in more than one piece of research. Therefore one is always left with the suspicion that worker A'sresults may be data specific and that were he to test them on workerB's date, they would not hold.