IR lessons for AI
Karen Spärck Jones · 1999
The claim that, on the basis of the author's experience, IR (information retrieval) has lessons for AI (artificial intelligence) is that its methods work in those situations where information demand, rather than supply, is underspecified, and the right response is to be broadly indicative rather than narrowly assertive. The right way of being indicative is then to allow many small and individually ambiguous clues to combine and interact within whatever match between user input and file material can be found. Altogether there are many information management tasks, from finding, to sorting, to reminding, that arise in many different working contexts, that are quite crude, and that can generally be done (appropriately and not just for want of anything better), in a sufficient-to-the-day mode because they are fundamentally inexact. Using weak, multiple, redundant cues is not a novel idea in AI; but the argument here is that they may be appropriate in far more natural language cases than originally envisaged, and also in ones where, even if the representation language used is not `ordinary' natural language it shares some important properties with it, for instance in search operations on the data represented by discourse in formal languages, as in software. There is a challenge to the IR methods described in their own world, namely that of terabyte files and minimalist users. But while how far they can be pushed in that case has clearly to be investigated, this does not imply that they should not be pushed as far as they will go in the other areas mentioned, i.e. in `real' AI. (3 pages)