lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Chuck Williams" <ch...@manawiz.com>
Subject RE: Relevance and ranking ...
Date Fri, 17 Dec 2004 19:32:29 GMT
Another issue will likely be the tf() and idf() computations.  I have a
similar desired relevance ranking and was not getting what I wanted due
to the idf() term dominating the score.  Lucene squares the contribution
of this term, which is not considered best practice in IR.  To address
these issues, I increased the base of the log for both tf() and idf()
(tones them down) and took a final square root on idf().  FYI, here are
the definitions I'm using for these methods -- similar definitions
should give you the ordering you want.  You might want to adjust
lengthNorm if you really want it to be linear (square root by default).
You should not have to touch coord().

    public float tf(float freq) {
        return 1.0f + (float)Math.log10(freq);
    }
    
    public float idf(int docFreq, int numDocs) {
      return (float)Math.sqrt(1.0 +
                              Math.log10(numDocs/(double)(docFreq+1)));
    }
    

Chuck

  > -----Original Message-----
  > From: Erik Hatcher [mailto:erik@ehatchersolutions.com]
  > Sent: Friday, December 17, 2004 4:06 AM
  > To: Lucene Users List
  > Subject: Re: Relevance and ranking ...
  > 
  > 
  > On Dec 17, 2004, at 6:09 AM, Gururaja H wrote:
  > > Thanks for the reply.  Is there any sample code which tells me how
to
  > > change these
  > > coord() factor,  overlapping, lenght normalizaiton etc.. ??
  > > If there are any please provide me.
  > 
  > Have a look at Lucene's DefaultSimilarity code itself.  Use that as
a
  > starting point - in fact you should subclass it and only override
the
  > one or two methods you want to tweak.
  > 
  > There are probably some other examples in Lucene's test cases, or
that
  > have been posted to the list but I don't have handy pointers to
them.
  > 
  > 	Erik
  > 
  > 
  > >
  > > Thanks,
  > > Gururaja
  > >
  > >
  > > Erik Hatcher <erik@ehatchersolutions.com> wrote:
  > > The coord() factor of Similarity is what controls a muliplier
factor
  > > for overlapping query terms in a document. The DefaultSimilarity
  > > already contains factors that allow documents with overlapping
terms
  > to
  > > get boosted. Is this not working for you? You may also need to
adjust
  > > length normalization factors. Check the javadocs on Similarity for
  > > details on implementing your own formulas. Also become familiar
with
  > > IndexSearcher.explain() and the Explanation so that you can see
how
  > > adjusting things affects the details.
  > >
  > > Erik
  > >
  > > On Dec 17, 2004, at 3:42 AM, Gururaja H wrote:
  > >
  > >> Hi,
  > >>
  > >> How to implement the following ? Please provide inputs ....
  > >>
  > >>
  > >> For example, if the search query has 5 terms (ibm, risc, tape,
drive,
  > >> manual) and there are 4 matching documents with the following
  > >> attributes, then the order should be as described below.
  > >>
  > >> Doc#1: contains terms (ibm, drive) and has a total of 100 terms
in
  > the
  > >> document.
  > >>
  > >> Doc#2: contains terms (ibm, risc, tape, drive) and has a total of
30
  > >> terms in the document.
  > >>
  > >> Doc#3: contains terms (ibm, risc, tape, drive) and has a total of
100
  > >> terms in the document.
  > >>
  > >> Doc#4: contains terms (ibm, risc, tape, drive, manual) and has a
  > total
  > >> of 300 terms in the document
  > >>
  > >> The search results should include all three documents since each
has
  > >> one or more of the search terms, however, the order should be
  > returned
  > >> as:
  > >>
  > >> Doc#4
  > >>
  > >> Doc#2
  > >>
  > >> Doc#3
  > >>
  > >> Doc#1
  > >>
  > >> Doc#4 should be first, since of the 5 search terms, it contains
all 5.
  > >>
  > >> Doc#2 should be second, since it has 4 of the 5 search terms and
of
  > >> the number of terms in the document, its ratio is higher than
Doc#3
  > >> (4/30). Doc#3 has 4 of the 5 terms, but its ratio is 4/100.
  > >>
  > >> Doc#1 is last since it only has 2 of the 5 terms.
  > >>
  > >>
  > >> ----
  > >>
  > >> Thanks,
  > >> Gururaja
  > >>
  > >>
  > >> __________________________________________________
  > >> Do You Yahoo!?
  > >> Tired of spam? Yahoo! Mail has the best spam protection around
  > >> http://mail.yahoo.com
  > >
  > >
  > >
---------------------------------------------------------------------
  > > To unsubscribe, e-mail: lucene-user-unsubscribe@jakarta.apache.org
  > > For additional commands, e-mail:
lucene-user-help@jakarta.apache.org
  > >
  > >
  > >
  > > ---------------------------------
  > > Do you Yahoo!?
  > >  Send holiday email and support a worthy cause. Do good.
  > 
  > 
  >
---------------------------------------------------------------------
  > To unsubscribe, e-mail: lucene-user-unsubscribe@jakarta.apache.org
  > For additional commands, e-mail: lucene-user-help@jakarta.apache.org


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


Mime
View raw message