[jira] [Commented] (LUCENE-8025) compute avgdl correctly for DOCS_ONLY

["Robert Muir (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/LUCENE-8025?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16226139#comment-16226139 ] 

Robert Muir commented on LUCENE-8025:
-------------------------------------

I opened this as a separate issue because we may want to backport it. Relevance pretty much falls apart in this case today with these sims (except Classic which does not use a pivot, hence unchanged). Here are my results on the bengali collection (I had it handy from working with the analyzer):

||Sim||Baseline MAP||DOCS_ONLY MAP (master)||DOCS_ONLY MAP (patch)||
|Classic|0.2266|0.1231|0.1231|
|BM25|0.2947|0.0531|0.1390|
|I(ne)B2|0.3074|0.0534|0.1485|
|I(ne)B1|0.2848|0.0529|0.1248|
|PL2|0.2856|0.0148|0.1377|

I can dig up some other datasets to confirm.

> compute avgdl correctly for DOCS_ONLY
> -------------------------------------
>
>                 Key: LUCENE-8025
>                 URL: https://issues.apache.org/jira/browse/LUCENE-8025
>             Project: Lucene - Core
>          Issue Type: Bug
>            Reporter: Robert Muir
>
> Spinoff of LUCENE-8007:
> If you omit term frequencies, we should score as if all tf values were 1. This is the way it worked for e.g. ClassicSimilarity and you can understand how it degrades. 
> However for sims such as BM25, we bail out on computing avg doclength (and just return a bogus value of 1) today, screwing up stuff related to length normalization too, which is separate.
> Instead of a bogus value, we should substitute sumDocFreq for sumTotalTermFreq (all postings have freq of 1, since you omitted them).



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

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