next up previous contents
Next: Designing a System Up: Stating the Problem Previous: Definitions of Locality

Finding Versus Asking

To make the best use of today's search engines the user must first deduce good keywords to search on. However, coming up with good keywords without wasting a lot of time first searching webpages for the most appropriate keywords is a matter of luck, intelligence, and experience. If you're clever, if you spend a lot of time on the web, and if you know how the people who create the data you're currently searching for think, then your queries are going to be much more exact and efficient than average. In other words, you can only come up with good keywords when you already know more or less what's out there. Almost all of the search effort is placed on the user's shoulders and not on the computer's.

Ideally, a system like Known Space shouldn't be necessary. We should just be able to ask our computers, ``Get me the name of that Chinese spice that smells like parsnips'' rather than have to wander around some virtual Chinese grocery. The problem with asking, however, is that commercial search engines are non-contextual. Their answers often mingle a few interesting pages with many uninteresting ones. Extracting those few good pages out of a morass of irrelevant pages is work.

Lacking the intelligence to establish a good idea of the user's particular context, today's search engines can only rely on raw textual occurrences to answer queries. Further, they pay no attention to the context of previous queries from that user to better answer the current query. Finally, the answers they provide will likely become less and less useful as the web expands in volume and variety. Unless a user can afford a horde of extremely knowledgeable human beings--or design a supersmart search engine--mere asking doesn't necessarily give us what us want.

Locality gives context cues to the search engine for free: once you're in a Chinese grocery, then asking the same spice question there is more likely to get you what you want because there is more context for the system to detect and use. Given enough context, even a dumb agent can produce smart answers. Locality is also a good idea because it helps us answer negative questions. If you search the entire neighborhood dealing with a particular topic and only found junk then you would know that the sought for data is not out there and will quit trying to find it. Without context, you are left in doubt about whether the data isn't there versus whether it's there but the keywords you selected were inappropriate to pinpoint it.


next up previous contents
Next: Designing a System Up: Stating the Problem Previous: Definitions of Locality
Gregory J. E. Rawlins
1/13/1998