FYI, Mike wrote some evaluation stuff for Nutch a long time ago. I
found it in the Sourceforge Attic:
http://cvs.sourceforge.net/viewcvs.py/nutch/nutch/src/java/net/nutch/quality/Attic/
This worked by querying a set of search engines, those in:
http://cvs.sourceforge.net/viewcvs.py/nutch/nutch/engines/
The results of each engine is scored by how much they differ from all of
the other engines combined. The Kendall Tau distance is used to compare
rankings. Thus this is a good tool to find out how close Nutch is to
the quality of other engines, but it may not not be a good tool to make
Nutch better than other search engines.
In any case, it includes a system to scrape search results from other
engines, based on Apple's Sherlock search-engine descriptors. These
descriptors are also used by Mozilla:
http://mycroft.mozdev.org/deepdocs/quickstart.html
So there's a ready supply of up-to-date descriptions for most major
search engines. Many engines provide a skin specifically to simplify
parsing by these plugins.
The code that implemented Sherlock plugins in Nutch is at:
http://cvs.sourceforge.net/viewcvs.py/nutch/nutch/src/java/net/nutch/quality/dynamic/
Doug
Andrzej Bialecki wrote:
Hi,
I found this paper, more or less by accident:
"Scaling IR-System Evaluation using Term Relevance Sets"; Einat Amitay,
David Carmel, Ronny Lempel, Aya Soffer
http://einat.webir.org/SIGIR_2004_Trels_p10-amitay.pdf
It gives an interesting and rather simple framework for evaluating the
quality of search results.
Anybody interested in hacking together a component for Nutch and e.g.
for Google, to run this evaluation? ;)