[
https://issues.apache.org/jira/browse/LUCENE-7996?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16645052#comment-16645052
]
Adrien Grand commented on LUCENE-7996:
--------------------------------------
WAND and other optimizations were the reason why I opened this issue and moved
it forward, though Robert made a good point in his initial comment on this
issue that negative scores are not only an implementation problem for WAND and
other optimizations, it also hurts relevance.
Doug mentioned that preventing negative scores might be an issue for LTR, but
I've come to think that these optimizations are actually interesting for
learning to rank since simple models could leverage these optimizations. If
they use a reasonable number of features (using FeatureField) and combine
scores via a linear combination, then the resulting query could be a boolean
query that efficiently skips irrelevant documents.
> Should we require positive scores?
> ----------------------------------
>
> Key: LUCENE-7996
> URL: https://issues.apache.org/jira/browse/LUCENE-7996
> Project: Lucene - Core
> Issue Type: Wish
> Reporter: Adrien Grand
> Priority: Minor
> Fix For: master (8.0)
>
> Attachments: LUCENE-7996.patch, LUCENE-7996.patch, LUCENE-7996.patch
>
>
> Having worked on MAXSCORE recently, things would be simpler if we required
> that scores are positive. Practically, this would mean
> - forbidding/fixing similarities that may produce negative scores (we have
> some of them)
> - forbidding things like negative boosts
> So I'd be curious to have opinions whether this would be a sane requirement
> or whether we need to be able to cope with negative scores eg. because some
> similarities that we want to support produce negative scores by design.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
