Re: score and frequency

Thu, 24 Jun 2004 06:34:56 -0700

Forget about the PhraseQuery, I'm stupid, it can't work like that. Because the phrase query will boost the documents which contain the search and not the documents which match exactly the search. So, the exact matches will come down. :-/

You need to have some information about the lucene documents to know if it's an exact match. Such as the number of terms in the documents. The problem is that this number is store in the lengthNorm and as it's encoded on 1 byte, you can't have it precisely. So, you should shunt the problem.
Here's another suggestion (a good one I hope):
Add another field containing the title as a Keyword. Then you just have to replace the PhraseQuery I told you to use by a TermQuery searching for the term (newField,search)
Of course, it will be a bit too restrictive to store the title without any transformation. You can for example store in this field the concatenation of the token given by your analyzer. Just don't forget to do the same transformation also for the search.
Sorry for the previous posts.

Franck

Niraj Alok wrote: 
Hi Franck,

You seem to be a genius in lucene !

I have done finally all that which you have suggested, but now when I am
searching for "lion", those terms are coming much below in terms of scores.

This is despite me setting the boost for the phrase query. Infact, this is
resulting in almost all the exact matches to come down.


Regards,
Niraj
----- Original Message -----
From: "Brisbart Franck" <[EMAIL PROTECTED]>
To: "Lucene Users List" <[EMAIL PROTECTED]>
Sent: Thursday, June 24, 2004 6:02 PM
Subject: Re: score and frequency



It may come from the boolean clauses. If you add your sub-queries with a
'required' flag, you'll only get the results matching all the words in
your query.
It can also come from the score which is different. If you set up a
threshold to return the results, it can be the problem.

Franck


Niraj Alok wrote:

Hi Franck,

Thank you so much for the detailed explanation.
However, when I tried to break up my MultiFieldQueryParser into a series


of

BooleanQueries, the result set has got reduced drastically.
Any idea why this could be happening?

Regards,
Niraj
----- Original Message -----
From: "Brisbart Franck" <[EMAIL PROTECTED]>
To: "Lucene Users List" <[EMAIL PROTECTED]>
Sent: Thursday, June 24, 2004 2:54 PM
Subject: Re: score and frequency




The MultiFieldQueryParser give you a BooleanQuery containing 1 query for
each field.
Something like:
          BooleanQuery
          /   |   |   \
        QF1  QF2 QF3  QF4    (QFx=Query for field x)

You can still use the MultiFieldQueryParser and create a BooleanQuery to
encapsulate the one parsed + the PhraseQuery, ie:
           BooleanQuery(created by you)
            /       \
          BQ      PhraseQuery

Or create the whole query (I think you should do that) and have
something like that:
           _BooleanQuery__
          /   |   |   \   \
        QF1  QF2 QF3  QF4  PhraseQuery      (QFx=Query for field x)

It's like parsing the following query:
(field1:query) (field2:query) (field3:query)...(fieldx:query)
(title:"query")~boost


Franck


Niraj Alok wrote:


I asked the previous question since I do not know how to use 

PhraseQuery

I have one booleanquery and one query.
The query is Query query =  MultiFieldQueryParser.parse( qs, searchLoc,
flags, new StandardAnalyzer(stop));

where qs is the word to be searched upon and searchLoc contains all the 

four


fields.

How do I insert a PhraseQuery here for title field only, and that too 

with


its boosted value?


Regards,
Niraj
----- Original Message -----
From: "Niraj Alok" <[EMAIL PROTECTED]>
To: "Lucene Users List" <[EMAIL PROTECTED]>
Sent: Thursday, June 24, 2004 2:00 PM
Subject: Re: score and frequency





Does it mean that I would need to abandon MultiFieldQueryParser?

Regards,
Niraj
----- Original Message -----
From: "Brisbart Franck" <[EMAIL PROTECTED]>
To: "Lucene Users List" <[EMAIL PROTECTED]>
Sent: Thursday, June 24, 2004 1:22 PM
Subject: Re: score and frequency





Hi,
first, what do you consider as an 'exact matching' ? It seems that


you

treat the search word by word, so 'lion sea' will be an 'exact match' 

of


'sea-lion'.
I think you should add a PhraseQuery to your query containing the


title

and with a big boost. So, you don't need to boost your title field. 

Only


the results matching exactly (for the PhraseQuery) will be boosted.

Franck


Niraj Alok wrote:



Hi Guys,

I seem to have run into rough weather again.
To describe the problem as concisely as possible, I have four fields


to



search upon : title , first para, rest of the paras and content (equal 

to


title + first para + rest of the para) .  I am doing this by using
MultiFieldQueryParser.



Now there is a very complicated ranking algrorithm specified by the

client and I have met most of them except one or two and really need 

your


help as all my other efforts have failed.



The most important rule is that exact matching titles should come 

first



, i.e. get higher scores.



I have given the highest boost factor to the title than the rest but 

the



problem comes up when there is some other title which has got just one 

word



matching. For e.g., if I search for lion, there is a title sea-lion 

which


also has the same boost factor as that of "lion" in the index. Also,
sea-lion has got some more "lion" in its first para or rest of the


paras

etc. such that its score comes higher than "lion".



Is there some way to get the exact matching titles higher scores?
Please reply soon.


Regards,
Niraj


----- Original Message -----
From: "Brisbart Franck" <[EMAIL PROTECTED]>
To: "Lucene Users List" <[EMAIL PROTECTED]>
Sent: Monday, June 07, 2004 12:50 PM
Subject: Re: score and frequency






It seems that you don't the length norm to be used. It's a factor 

which



normalize the score of a doc depending on the size of the searched 

field



of the doc. It's the field which make that 'ground ice' has a 

higher

score than 'ice hockey: British Sekonda Superleague Play-Off
Championship: finals' because it only has 2 terms.
So, I suggest you to override the lengthNorm method and to ignore


the

numTokens parameter.
NB: The length norm is computed during the indexation and the norm


are


store in the index (in the _aaa.f# files). So, you need to do 

re-index


your data, and use this similarity during the indexation.

Cheers,
Franck


Niraj Alok wrote:




I have set the searcher.setSimilarity as well as also tried 

setting

the



coord factor to 1.

The problem as given by an example is : Lets say I have titles to 

be

displayed depending upon the search.
E.g if i have "ice hockey" as the search item and if it is default
similarity, my results are :

ice hockey0.99999994
ice hockey0.75
ice hockey0.75
winter Olympics: hockey, ice, medallists0.17402513
ice age0.073680125
National Hockey League0.020266924
Cracking the Ice Age0.018420031
ground-ice0.011512519
ice hockey: British Sekonda Superleague Play-Off Championship:
finals0.0069075115
(the numbers indicating the score).


But if i set the similarity as my overridden one, the results 

become:


ice hockey0.99999994
ice hockey0.75
ice hockey0.75
ice age0.22104037
winter Olympics: hockey, ice, medallists0.17402513
National Hockey League0.060800765
Cracking the Ice Age0.055260092
ground-ice0.034537554
ice hockey: British Sekonda Superleague Play-Off Championship:
finals0.020722535


I want all the titles which have both "ice" and "hockey" to come 

above



the



rest (to have higher scores)
Meaning i would wish the results to appear like:

ice hockey
ice hockey
ice hockey
winter Olympics: hockey, ice, medallists
ice hockey: British Sekonda Superleague Play-Off Championship:


finals


ice age
National Hockey League
Cracking the Ice Age
ground-ice

My overriden similarity class contains just this method:
public float coord(int overlap, int maxOverlap) {

return 1.0f;

}





I feel it is the weight factor which is producing indesirable 

results.



Any



help in this regard would be highly appreciated.

Regards,
Niraj

----- Original Message -----
From: "Brisbart Franck" <[EMAIL PROTECTED]>
To: "Lucene Users List" <[EMAIL PROTECTED]>
Sent: Friday, June 04, 2004 8:46 PM
Subject: Re: score and frequency







Hi,

Be careful to set the default similarity
'Similarity.setDefault(similarity)' before creating your search


instance



(IndexSearcher).
If you change the default similarity after, you'll still use the


old


one.



You'd better use the 'searcher.setSimilarity' method on your 

searcher.



Franck


Phil brunet wrote:





Hi to all.

Maybe the term frequency is not the only parameter you need to 

override



to "customize" the score attributed by Lucene.

Maybe you should consider the normalisation factor, the idf and 

the


coord factor ?

Philippe






From: "Niraj Alok" <[EMAIL PROTECTED]>
Reply-To: "Lucene Users List" <[EMAIL PROTECTED]>
To: "Lucene Users List" <[EMAIL PROTECTED]>
Subject: Re: score and frequency
Date: Fri, 4 Jun 2004 15:13:32 +0530

Hi Erik,

Thanks for the suggestion.

I tried this:
public class RelevanceSimilarity extends DefaultSimilarity

{

public float tf(float freq) {

System.out.println("discounting frequency");

return (float)1;

}

}



and in my query class, I used :

Similarity.setDefault(similarity);

Hits hits = is.search(query);

for(i = 0; i < hits.length(); i ++)

result = result + hits.score(i);



However, this is still not giving me the expected result. Do I 

need



to



do





something else?


Regards,
Niraj

----- Original Message -----
From: "Erik Hatcher" <[EMAIL PROTECTED]>
To: "Lucene Users List" <[EMAIL PROTECTED]>
Sent: Friday, June 04, 2004 1:55 PM
Subject: Re: score and frequency







On Jun 4, 2004, at 2:52 AM, Niraj Alok wrote:





Hi,

I am having some problems with the score of lucene.
I am trying to get the results displayed according to


hits.score


and





it is giving the results correctly.
However I do not want the frequency factor to be used for the
computation of the score.

Is it possible to get the score which does not have the 

frequency



factor in it ?

Have a look at the javadocs for Similarity. DefaultSimilarity 

is


used





unless otherwise specified. You could subclass that and 

override


this:





public float tf(float freq) {
return (float)Math.sqrt(freq);
}

and return 1.0.  This might give you the effect you want.

Erik



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

-

-


To unsubscribe, e-mail: 

[EMAIL PROTECTED]



For additional commands, e-mail: 

[EMAIL PROTECTED]




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

-

To unsubscribe, e-mail: 

[EMAIL PROTECTED]


For additional commands, e-mail: 

[EMAIL PROTECTED]



_________________________________________________________________

Bloquez les fen�tres pop-up, c'est gratuit ! 

http://toolbar.msn.fr


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

To unsubscribe, e-mail: 

[EMAIL PROTECTED]

For additional commands, e-mail: 

[EMAIL PROTECTED]



--
Franck Brisbart
R&D
http://www.kelkoo.com



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


To unsubscribe, e-mail: 

[EMAIL PROTECTED]

For additional commands, e-mail: 

[EMAIL PROTECTED]





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

To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail:


[EMAIL PROTECTED]


--
Franck Brisbart
R&D
http://www.kelkoo.com



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

To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail:


[EMAIL PROTECTED]



--
Franck Brisbart
R&D
http://www.kelkoo.com


---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]





---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]






---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]



--
Franck Brisbart
R&D
http://www.kelkoo.com


---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]




--
Franck Brisbart
R&D
http://www.kelkoo.com


---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]





--
Franck Brisbart
R&D
http://www.kelkoo.com


---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]

Reply via email to