lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Andrzej Bialecki ...@getopt.org>
Subject Solved (Re: Document visible by Term, but not search)
Date Fri, 26 Aug 2005 17:33:02 GMT
Hi list,

This is just to let you know that I found the reason (Dan sent me a 
small sample index off-list), and I thought that the reason for this 
error was obscure and tricky enough that you might be interested in the 
solution.

The problem lied in custom boost values. It was impossible to find the 
documents using the high-level search() interface. If you remember, this 
interface skips the lowest-scoring hits, among others documents with 
score==0  :-)

How can the score be 0 if the document matches (and it matched, because 
it clearly contained the term from the query)? I implemented a version 
of HitCollector that collects all hits, in order to investigate this. 
Running a query "testField:test" against that sample index I got 1 hit 
with score 0, and this explanation:

     0.0000 fieldWeight(testField:test in 0), product of:
       1.0000 tf(termFreq(testField:test)=1)
       0.3069 idf(docFreq=1)
       0.0000 fieldNorm(field=testField, doc=0)

Under normal circumstances fieldNorm is never 0 ... unless a boosting 
has been applied. In this case the original poster didn't apply boost=0, 
but some other (small) value. Boost values are encoded floats with very 
coarse resolution. In this case this resulted in fieldNorm falling below 
resolution of the encoded float. The fractional part was lost in this 
case, because it was too small to be encoded, so that the fieldNorm 
became 0. As a consequence, the score became 0 too, even though the 
document matched ...

-- 
Best regards,
Andrzej Bialecki     <><
  ___. ___ ___ ___ _ _   __________________________________
[__ || __|__/|__||\/|  Information Retrieval, Semantic Web
___|||__||  \|  ||  |  Embedded Unix, System Integration
http://www.sigram.com  Contact: info at sigram dot com


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