I think the Reuters corpus is pretty good and it pretty well known in
the community. Probably the most important part would be to build up
a set of judgments. I don't think it is too hard to come up w/
50-100 questions/queries, but creating the relevance pool will be
more difficult. I suppose we could setup a social networking site to
harvest judgments... :-)
The 4M queries would be good for load testing.
Wikipedia stuff is good, but you need to be able to handle/remove the
redirects, otherwise you have a tendency to get redirect pages as
your top matches due to length normalization. Plus it is really big
to download.
On Mar 20, 2007, at 6:58 AM, Karl Wettin (JIRA) wrote:
[ https://issues.apache.org/jira/browse/LUCENE-836?
page=com.atlassian.jira.plugin.system.issuetabpanels:comment-
tabpanel#action_12482367 ]
Karl Wettin commented on LUCENE-836:
------------------------------------
Regarding data and user queries, I have a 150 000 document corpus
with 4 000 000 queries that I might be able to convince the owners
to release. It is great data, but a bit politically incorrect
(torrents).
There is some simple Wikipedia harvesting in LUCENE-826, and I'm in
the middle of rewriting it to a more general Wikipedia library for
text mining purposes. Perhaps you have some ideas you want to put
in there? I plan something like this:
public class WikipediaCorpus {
Map<String, String> wikipediaDomainPrefixByLanguageISO
Map<URL, WikipediaArticle> harvestedArticle
public WikipediaArticle getArticle(String languageISO, String
title) {
..
}
}
public class WikipediaArticle {
WikipediaArticle(URL url) {
..
}
String languageISO;
String title;
String[] contentParagraphs
Date[] modified;
Map<String, String> articleInOtherLanguagesByLanguageISO
}
Benchmarks Enhancements (precision/recall, TREC, Wikipedia)
-----------------------------------------------------------
Key: LUCENE-836
URL: https://issues.apache.org/jira/browse/LUCENE-836
Project: Lucene - Java
Issue Type: New Feature
Components: Other
Reporter: Grant Ingersoll
Priority: Minor
Would be great if the benchmark contrib had a way of providing
precision/recall benchmark information ala TREC. I don't know
what the copyright issues are for the TREC queries/data (I think
the queries are available, but not sure about the data), so not
sure if the is even feasible, but I could imagine we could at
least incorporate support for it for those who have access to the
data. It has been a long time since I have participated in TREC,
so perhaps someone more familiar w/ the latest can fill in the
blanks here.
Another option is to ask for volunteers to create queries and make
judgments for the Reuters data, but that is a bit more complex and
probably not necessary. Even so, an Apache licensed set of
benchmarks may be useful for the community as a whole. Hmmm....
Wikipedia might be another option instead of Reuters to setup as a
download for benchmarking, as it is quite large and I believe the
licensing terms are quite amenable. Having a larger collection
would be good for stressing Lucene more and would give many users
a demonstration of how Lucene handles large collections.
At any rate, this kind of information could be useful for people
looking at different indexing schemes, formats, payloads and
different query strategies.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
------------------------------------------------------
Grant Ingersoll
http://www.grantingersoll.com/
http://lucene.grantingersoll.com
http://www.paperoftheweek.com/
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
