lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ahmet Arslan <iori...@yahoo.com.INVALID>
Subject Re: Learning to Rank algorithms in Lucene
Date Tue, 18 Aug 2015 12:58:42 GMT
Hi Ajinkya,

I don't think there exists any production-ready LtR-Lucene/Solr setup.

LtR simply re-rank top N (typically 1000) documents. 
Fetching top N documents is what we do today with Lucene.

There is an API for re-rank in Lucene/Solr but no LtR support yet.
https://cwiki.apache.org/confluence/display/solr/Query+Re-Ranking

Here are the difficulties/problems :

* LtR requires training data (probably labelled by humans)
* It is hard to decide the feature set. Also it differs from system to system.
* Query-dependeny features must be calculated for the top N documents at query/retrieval time,
which may be slow.

Today, generally function queries are used to combine recency, popularity, star, product/document
quality, price, etc into scoring function.
This approach is unsupervised therefore requires no training data.

Ahmet



On Tuesday, August 18, 2015 10:34 AM, Ajinkya Kale <kaleajinkya@gmail.com> wrote:
Are there any existing packages/examples or prior experience on using
Learning to Rank (or Machine Learned Ranking) algorithms as custom
Scorer/Ranker for lucene or solr ?
How do people deploy Learning to Rank models with Lucene backends ?

--ajinkya

---------------------------------------------------------------------
To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-user-help@lucene.apache.org


Mime
View raw message