Hello,

I'm helping out a student interested in using query and click logs to build
custom relevance models for Lucene.  Step #1 is finding a good dataset that
contains the needed data.  I've looked around, found a few things, but nothing
that looks very good.

I was wondering if anyone has any dataset suggestions?

Thanks,
Otis
----
Sematext :: http://sematext.com/ :: Solr - Lucene - Nutch
Lucene ecosystem search :: http://search-lucene.com/

Hi Otis,
you may find some resources (mainly code, not datasets) for your student's
use case at [1] .
Also I know LWE from LucidImagination [2] has a click & scoring framework
but that component is not open source at the moment; however I don't know if
they used also publicly available datasets to build such a feature.
My 0.2 cents,
Tommaso

[1] : http://code.google.com/p/oluolu
[2] :
http://www.lucidimagination.com/enterprise-search-solutions/lucidworks/1.6

2011/3/2 Otis Gospodnetic <[EMAIL PROTECTED]>

> Hello,
>
> I'm helping out a student interested in using query and click logs to build
> custom relevance models for Lucene.  Step #1 is finding a good dataset that
> contains the needed data.  I've looked around, found a few things, but
> nothing
> that looks very good.
>
> I was wondering if anyone has any dataset suggestions?
>
> Thanks,
> Otis
> ----
> Sematext :: http://sematext.com/ :: Solr - Lucene - Nutch
> Lucene ecosystem search :: http://search-lucene.com/
>
>

On 3/2/11 3:39 AM, Otis Gospodnetic wrote:
> Hello,
>
> I'm helping out a student interested in using query and click logs to build
> custom relevance models for Lucene.  Step #1 is finding a good dataset that
> contains the needed data.  I've looked around, found a few things, but nothing
> that looks very good.
>
> I was wondering if anyone has any dataset suggestions?

The (in)famous AOL dataset comes to my mind, and it's very good, maybe
even too good :) AOL officially pulled it back, but it's still available
and IMHO legitimate to use - it was a blunder all right but it carried a
suitable license and things can't be un-published ...

--
Best regards,
Andrzej Bialecki     <><
  ___. ___ ___ ___ _ _   __________________________________
[__ || __|__/|__||\/|  Information Retrieval, Semantic Web
___|||__||  \|  ||  |  Embedded Unix, System Integration
http://www.sigram.com  Contact: info at sigram dot com

Hi Andrzej,

Thanks for bringing up that AOL dataset (I've got a copy of that stashed away),
because the person I'm helping looked at this, and we thought it didn't have all
the data one needs to build custom relevance models.
Here is a small sample:

 AnonID     Query   QueryTime       ItemRank        ClickURL
 217        lottery 2006-03-01 11:58:51     1       http://www.calottery.com
 217        lottery 2006-03-01 11:58:51     1       http://www.calottery.com
 217        ameriprise.com  2006-03-01 14:06:23     1    
  http://www.ameriprise.com
 217        susheme 2006-03-02 12:31:08
 217        united.com      2006-03-03 14:54:13
 217        mizuno.com      2006-03-07 22:41:17     1    
  http://www.mizuno.com
 217        p; .; p;' p; ' ;' ;';   2006-03-09 12:09:27
 217        p; .; p;' p; ' ;' ;';   2006-03-09 12:09:35
 217        buddylis        2006-03-16 15:23:33
 217        bestasiancompany.com    2006-03-20 15:15:43     1    
  http://www.bestasiancompany.com
 217        lottery 2006-03-27 14:10:38     1       http://www.calottery.com
 217        lottery 2006-03-27 16:34:59     1       http://www.calottery.com
 217        ask.com 2006-03-31 14:31:10     1       http://www.ask.com

For instance, in order to  build custom relevance models, wouldn't we need to
have the actual corpus/index associated with this data in order to get the base
relevance scores first?

Or could one just look at clicks where ItemRank is low (meaning they were not
close to the top of search results) and apply some algo that essentially
produces a boost score that stands on its own and is applied on top of the
relevance score at search time?
Would it make sense to have a global boost score for each document, or would
that need to be query-specific and thus applied at query-time and not at
index-time?

If you have an idea how one could/should go about using just the above to  build
custom relevance models for Lucene, I'm all eyeballs.

Thanks,
Otis
----
Sematext :: http://sematext.com/ :: Solr - Lucene - Nutch
Lucene ecosystem search :: http://search-lucene.com/

----- Original Message ----
> From: Andrzej Bialecki <[EMAIL PROTECTED]>
> To: [EMAIL PROTECTED]
> Sent: Wed, March 2, 2011 5:07:55 AM
> Subject: Re: Query & click logs for custom Lucene relevance models
>
> On 3/2/11 3:39 AM, Otis Gospodnetic wrote:
> > Hello,
> >
> > I'm  helping out a student interested in using query and click logs to build
> >  custom relevance models for Lucene.  Step #1 is finding a good dataset  
that
> > contains the needed data.  I've looked around, found a few  things, but
>nothing
> > that looks very good.
> >
> > I was  wondering if anyone has any dataset suggestions?
>
> The (in)famous AOL  dataset comes to my mind, and it's very good, maybe even
>too good :) AOL  officially pulled it back, but it's still available and IMHO
>legitimate to use -  it was a blunder all right but it carried a suitable
>license and things can't be  un-published ...
>
> -- Best regards,
> Andrzej Bialecki      <><
>  ___. ___ ___ ___ _ _    __________________________________
> [__ || __|__/|__||\/|  Information  Retrieval, Semantic Web
> ___|||__||  \|  ||  |  Embedded  Unix, System Integration
> http://www.sigram.com  Contact: info at sigram dot com
>
>