[
https://issues.apache.org/jira/browse/LUCENE-965?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12516185
]
Charlie Zhao commented on LUCENE-965:
-------------------------------------
Regarding the approach to compute avgDL, this patch goes like this:
+ float avgDL=0.0f;
+ for (int i=0; i<norms.length;i++) {
+ avgDL += normDecoder[norms[i] & 0xFF];
+ }
+ avgDL /= norms.length * 1.0f;
But may I suggest the alternative?
float CL = 0.0f;
float avgDL = 0.0f;
float aDL = 0.0f;
for (int i=0; i<norms.length;i++) {
aDL = 1.0f / normDecoder[norms[i] & 0xFF] ;
aDL *= aDL;
CL += aDL;
}
avgDL = CL / norms.length;
Let us see a toy example:
2 docs in index
|D| avgD |D|/avgD norm avgNorm
norm/avgNorm
D1 4 10 2/5 1/2 3/8 4/3
D2 16 10 8/5 1/4 3/8 2/3
norm/avgNorm is what we got from the patch code and D1>D2
|D|/avgD is what we got from the suggested alternative code and D1 < D2
They have totally flipped the relationship between D1 and D2.
My impression of the Axiomatic Retrieval Function is: it still tries to
penalize longer doc. So maybe the alternative code is what we need?
By the same token, |D| != Similarity.decodeNorm(fieldNorms[doc]).
Note: since we are recovering from the norm, so avgDL and DL != their original
absolute value. But they suffice for the scoring purpose.
Based on Doug's previous comment, I totally agree that avgDL should be
pre-computed and cached in the searcher before where the rubber meets the road.
And the cost might be invisible if we warm up the searcher first. Thanks for
explaining.
Not sure where Doron implemented 1 / sqrt((1 - Slope) * Pivot + (Slope) *
Doclen). Since LUCENE-836 looks will be committed soon. I am really excited to
see which similarity function will prevail in this era.
BTW, anyone would like to share how to read Lucene patches more efficiently? I
mean I had hard time to make sense of those +s and -s independently from their
source files. Is there a way to plug in a patch into my local source
repository, so I can diff with my favorite diff tool? Thanks in advance.
> Implement a state-of-the-art retrieval function in Lucene
> ---------------------------------------------------------
>
> Key: LUCENE-965
> URL: https://issues.apache.org/jira/browse/LUCENE-965
> Project: Lucene - Java
> Issue Type: Improvement
> Components: Search
> Affects Versions: 2.2
> Reporter: Hui Fang
> Attachments: axiomaticFunction.patch
>
>
> We implemented the axiomatic retrieval function, which is a state-of-the-art
> retrieval function, to
> replace the default similarity function in Lucene. We compared the
> performance of these two functions and reported the results at
> http://sifaka.cs.uiuc.edu/hfang/lucene/Lucene_exp.pdf.
> The report shows that the performance of the axiomatic retrieval function is
> much better than the default function. The axiomatic retrieval function is
> able to find more relevant documents and users can see more relevant
> documents in the top-ranked documents. Incorporating such a state-of-the-art
> retrieval function could improve the search performance of all the
> applications which were built upon Lucene.
> Most changes related to the implementation are made in AXSimilarity,
> TermScorer and TermQuery.java. However, many test cases are hand coded to
> test whether the implementation of the default function is correct. Thus, I
> also made the modification to many test files to make the new retrieval
> function pass those cases. In fact, we found that some old test cases are not
> reasonable. For example, in the testQueries02 of TestBoolean2.java,
> the query is "+w3 xx", and we have two documents "w1 xx w2 yy w3" and "w1 w3
> xx w2 yy w3".
> The second document should be more relevant than the first one, because it
> has more
> occurrences of the query term "w3". But the original test case would require
> us to rank
> the first document higher than the second one, which is not reasonable.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
