[
https://issues.apache.org/jira/browse/LUCENE-3364?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13080439#comment-13080439
]
John Wang commented on LUCENE-3364:
-----------------------------------
Hi Yonik:
In Collector, the decision time to cut of a score would be too late, e.g.
float score = Scorer.score(); <--- this is where the cost would occur.
boolean cutOff = decide(score)
In my example, my score impl:
float s1 = cheapCalc(docid);
float s2 = expensiveCalc(docid);
return s1+s2;
So now if I know expensiveCalc is bounded by N, and cheapCalc returns a very
small number, I can simply skip s2 calculation because this doc would be thrown
out anyway.
Hope I am making sense :)
-John
> Add score threshold into Scorer.score()
> ---------------------------------------
>
> Key: LUCENE-3364
> URL: https://issues.apache.org/jira/browse/LUCENE-3364
> Project: Lucene - Java
> Issue Type: Improvement
> Components: core/query/scoring
> Reporter: John Wang
>
> This is an optimization for scoring.
> Given a Scorer.score() implementation, where features are gathered to
> calculate a score.
> Proposal, add a parameter to score, e.g. score(float threshold)
> This threshold is the minimum score to "beat" to make it to the current
> PriorityQueue. This could potential save a great deal of wasted calculation
> in the cases where recall is large.
> In our case specifically, some of the features needed to do calculation can
> be expensive to obtain, it would be nice to have a place to decide whether or
> not even fetching these features are necessary.
> Also, if we know the score would be low, simply threshold can be returned.
> Let me know if this makes sense and I can work on a patch.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
