“Real world” searching panel at SIGIR 97
Michael Lesk · ACM SIGIR Forum · 1998
Reported by Michael LeskThe SIGIR panel session on "Real World IR" emphasized content, practicality, and the lack of relevance of most SIGIR papers to the search business.Doug Cutting, Jan Pedersen, Terry Noreault and Matt Koll participated.Doug Cutting works at Excite, a web search engine which indexes 50 million pages, using 50 GB of disk space for its index.Their peak load is hundreds of queries per second, and their revenue comes from advertising at about 2 cents per search.About 5% of their gross revenue is available for search hardware.Once peak-load and redundancy needs are factored in, only about a hundredth of a cent per search remain.This means that the hardware to search the fifty million pages at one second per query cannot cost more than $6,000.Their great need is for fast throughput -they can meet their cost objective only by using cheap features.Thus, they have term suggestions, but not clustering.Many very practical considerations intrude.Users try to fool the search engines by stuff'mg their pages, i.e. adding words believed to be frequently searched for in ways which will not display (in META tags, in background color, or small type).Stuffed pages must be detected and penalized.Duplicate pages are also detected and removed.About 80% of users look only at the first page.Thus, precision is much more important than recall, and stemming is not useful.Queries are very short.Last year, they averaged 1.5 words, but are now up to over 2 words per query.Interestingly, in the European language servers query lengths are at 1.5 words; users there have not yet learned enough about the search engines.Another example of user learning is that the use of quoted strings doubled this January when phrase searching was enabled (last year a search for "laptop computer" was interpreted as "laptop AND computer"); quoted strings are now in 10% of searches instead of 5% even though there was no announcement of the new feature.People do like term suggestions.