lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Frédéric Glorieux <emp...@fictif.org>
Subject Re: Can I use DFISimilarity for search on an an index written with BM25Similarity ?
Date Tue, 18 Jun 2019 11:07:04 GMT
Confirmed.

It is not a full lucene app and it is in French, but this query is very 
good to observe the size bias in scoring of unequal docs (chapters of books)

http://obvil.lip6.fr/alix/snip.jsp?q=dédicace&sort=lmd

BM25 (or LMD) are incredibly robust to document size.

This can change deeply the relevance of MoreLikeThis queries and other 
semantic calculations, great.

--
Frédéric


Le 14/06/2019 à 18:30, Adrien Grand a écrit :
> Yes, you can use DFISimilarity with an index constructed with
> BM25Similarity. No need to reindex.
>
> On Fri, Jun 14, 2019 at 1:05 PM Frédéric Glorieux <emploi@fictif.org> wrote:
>> Hi,
>>
>> I'm working on literature texts (French).
>>
>> My users are interested in relevance tweaking to have the most suggested
>> texts (for their taste) in top results.
>>
>> Change similarity at query time is less expensive than reindex all.
>>
>> I checked that BM25 needs to write “norms“ to keep document length.
>>
>> Have I missed something ? DFISimilarity seems to write and use norms
>> from SimilarityBase, where it is written
>>
>> computeNorms  «Encodes the document length in the same way as {@link
>> BM25Similarity}»
>>
>> For my first experiences, it seems that results with DFISimilarity at
>> query time are the same with an index encoded with default
>> BM25Similarity or DFI.
>>
>> Can some gurus confirm with their experience ?
>>
>> Thanks in advance (and lucene is really a good piece of software).
>>
>> --
>> Frédéric
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
>> For additional commands, e-mail: java-user-help@lucene.apache.org
>>
>

---------------------------------------------------------------------
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