Boosting Vs Sorting

Boosting Vs Sorting

	Rakesh Shete
	Dec 21, 2007 at 3:51 am

Hi all,

I am using Hibernate Search (http://www.hibernate.org/410.html) which is a wrapper around Lucene for performing search over info stored in the DB. I have questions related to Lucene boosting Vs sorting:

Is index time boosting of documents and fields better than specifying sorting parameters at search time?

I have been browsing through the Lucene mail archives for an answer to this. Going through them and reading on stuff related to Lucene scoring, my understanding is that if I know upfront at index time that the relevance order of results is based on certain fields, then, it is better to have index time boosting of documents and fields. Am I right here?

My requirements are like:
Results having an exact match to the input query string should have highest preference followed by an exact match with field1, field2, field3 and then followed by search query substring (or near match) match with field1, field2, field3.

Any suggestions are most welcome.

--Rakesh S

_________________________________________________________________
Post free property ads on Yello Classifieds now! www.yello.in
http://ss1.richmedia.in/recurl.asp?pid=219

Search Discussions

6 responses
Oldest
Nested

Erick Erickson

at Dec 21, 2007 at 2:46 pm

⇧

From my perspective, index-time boosting and sorting are apples

and oranges.

According to a post from Hoss, index-time boosting is a way of
saying that "Field x in this document is more important than
field x in other documents". Query-time boosts are a way of
saying "I care about field X more than field Y across *all*
documents".

So index time boosting doesn't seem to relate to your problem since
you really want to compare field x across all documents. It seems
that query-time boosting is more relevant.

So, leaving aside how you form your "similar" q

On Dec 20, 2007 10:50 PM, Rakesh Shete wrote:

Hi all,

I am using Hibernate Search (http://www.hibernate.org/410.html) which is a
wrapper around Lucene for performing search over info stored in the DB. I
have questions related to Lucene boosting Vs sorting:

Is index time boosting of documents and fields better than specifying
sorting parameters at search time?

I have been browsing through the Lucene mail archives for an answer to
this. Going through them and reading on stuff related to Lucene scoring, my
understanding is that if I know upfront at index time that the relevance
order of results is based on certain fields, then, it is better to have
index time boosting of documents and fields. Am I right here?

My requirements are like:
Results having an exact match to the input query string should have
highest preference followed by an exact match with field1, field2, field3
and then followed by search query substring (or near match) match with
field1, field2, field3.

Any suggestions are most welcome.

--Rakesh S

_________________________________________________________________
Post free property ads on Yello Classifieds now! www.yello.in
http://ss1.richmedia.in/recurl.asp?pid=219

reply | permalink

Erick Erickson

at Dec 21, 2007 at 2:53 pm

⇧

OK, I'm trying to adjust to a Mac and my keyboard shortcuts sometimes
lead me to send the mail when I didn't intend. Sorry about that...

So, leaving aside how you form your "similar" query, I *think* you
want to form two clauses, your "exact" and your "similar" and
boost them individually, combined in a boolean query.

This will still interleave the results I think. But it's also a valid
question whether this is good or bad. Is it *really* better for your
users to see a low-relevance query that happens to have the exact
words in it before a very-high ranking but not quite exact response?
That, of course it up to your product manager....

If it is really a requirement, it seems to me that you would be able to
just form the two queries independently, then just post-process them.
One query is the exact version, and the second query is the similar one.
Then just combine the results as you please by iterating the hits
object for the exact query then following it by the same for the similar.

I don't see how sorting relates to your problem at all....

Best
Erick

On Dec 21, 2007 9:46 AM, Erick Erickson wrote:

From my perspective, index-time boosting and sorting are apples
and oranges.

According to a post from Hoss, index-time boosting is a way of
saying that "Field x in this document is more important than
field x in other documents". Query-time boosts are a way of
saying "I care about field X more than field Y across *all*
documents".

So index time boosting doesn't seem to relate to your problem since
you really want to compare field x across all documents. It seems
that query-time boosting is more relevant.

So, leaving aside how you form your "similar" q

On Dec 20, 2007 10:50 PM, Rakesh Shete wrote:

Hi all,

I am using Hibernate Search (http://www.hibernate.org/410.html) which is
a wrapper around Lucene for performing search over info stored in the DB. I
have questions related to Lucene boosting Vs sorting:

Is index time boosting of documents and fields better than specifying
sorting parameters at search time?

I have been browsing through the Lucene mail archives for an answer to
this. Going through them and reading on stuff related to Lucene scoring, my
understanding is that if I know upfront at index time that the relevance
order of results is based on certain fields, then, it is better to have
index time boosting of documents and fields. Am I right here?

My requirements are like:
Results having an exact match to the input query string should have
highest preference followed by an exact match with field1, field2, field3
and then followed by search query substring (or near match) match with
field1, field2, field3.

Any suggestions are most welcome.

--Rakesh S

_________________________________________________________________
Post free property ads on Yello Classifieds now! www.yello.in
http://ss1.richmedia.in/recurl.asp?pid=219

reply | permalink

Rakesh Shete

at Dec 21, 2007 at 5:51 pm

⇧

Hi Eric,

I don't see how sorting relates to your problem at all....

Could you just explain how is sorting different from boosting?

I have been trying to figure this out. Going through "Lucene In Action" my
understanding of sorting is that it will kind of second level of ordering
after the query results have been scored (Not sure if the relevance established
by scoring is lost in this process).

Is it *really* better for your users to see a low-relevance query
that happens to have the exact words in it before a very-high
ranking but not quite exact response?

Nopes. Thats the last thing my product manager will want.

Lets take an example to simplify this:

I have fields like title, description, tags. Now when I search for a term
"Indoor Photography" then I would like the results with exact match in title to be
more important than in description or tags. However, if there is an exact match in description
then it should be given more preference than the partial match in title.

Going by the points mentioned below and as per one of your posts
(http://mail-archives.apache.org/mod_mbox/lucene-java-user/200609.mbox/%[email protected]%3E)
I understand that I need to specify query time boosting like this:

title:Indoor Photography^2.5 description:Indoor Photography^1.5 tags:
Indoor Photography^1.2

Let me know if this would help my cause.

Thnx for ur time n the valuable info.

--Rakesh S

Date: Fri, 21 Dec 2007 09:53:02 -0500
From: [email protected]
To: [email protected]
Subject: Re: Boosting Vs Sorting

OK, I'm trying to adjust to a Mac and my keyboard shortcuts sometimes
lead me to send the mail when I didn't intend. Sorry about that...

So, leaving aside how you form your "similar" query, I *think* you
want to form two clauses, your "exact" and your "similar" and
boost them individually, combined in a boolean query.

This will still interleave the results I think. But it's also a valid
question whether this is good or bad. Is it *really* better for your
users to see a low-relevance query that happens to have the exact
words in it before a very-high ranking but not quite exact response?
That, of course it up to your product manager....

If it is really a requirement, it seems to me that you would be able to
just form the two queries independently, then just post-process them.
One query is the exact version, and the second query is the similar one.
Then just combine the results as you please by iterating the hits
object for the exact query then following it by the same for the similar.

I don't see how sorting relates to your problem at all....

Best
Erick

On Dec 20, 2007 10:50 PM, Rakesh Shete wrote:

Hi all,

I am using Hibernate Search (http://www.hibernate.org/410.html) which is
a wrapper around Lucene for performing search over info stored in the DB. I
have questions related to Lucene boosting Vs sorting:

Is index time boosting of documents and fields better than specifying
sorting parameters at search time?

I have been browsing through the Lucene mail archives for an answer to
this. Going through them and reading on stuff related to Lucene scoring, my
understanding is that if I know upfront at index time that the relevance
order of results is based on certain fields, then, it is better to have
index time boosting of documents and fields. Am I right here?

My requirements are like:
Results having an exact match to the input query string should have
highest preference followed by an exact match with field1, field2, field3
and then followed by search query substring (or near match) match with
field1, field2, field3.

Any suggestions are most welcome.

--Rakesh S

_________________________________________________________________
Post free property ads on Yello Classifieds now! www.yello.in
http://ss1.richmedia.in/recurl.asp?pid=219

_________________________________________________________________
Post free property ads on Yello Classifieds now! www.yello.in
http://ss1.richmedia.in/recurl.asp?pid=219

reply | permalink

Erick Erickson

at Dec 21, 2007 at 10:16 pm

⇧

See below...

On Dec 21, 2007 12:50 PM, Rakesh Shete wrote:

Hi Eric,

I don't see how sorting relates to your problem at all....

I often think of sorting as being orthogonal to boosting. They're really
unrelated.
Boosting changes how the scoring of documents work. Sorting ignores scoring
and arranges the results lexically. You can *only* sort on fields that are a
single
token. I'm cheating a little here and you can implement your own
sorts, but that's another story.

Maybe this would help. Say you were indexing books and wanted the results
presented to the user by title. You could index a "titlesort" field that had
the
title lowercased and all spaces replaced with underscores. Then, you could
sort the result of all books containing "solar energy" by title. Where's the
score
here? The only relevance score has here is that no book in the result set
will
have a score of 0.

I did, at one point, have to sort by score then sub-sort by title. That is,
present the user with the top scoring documents sub-sorted by title.
This involved using relevancy as the primary sort and sub-sorting by
title. But the problem here is that scores of 0.98374 wouldn't be in the
same bucket as a score of 0.98375. Search the mail archive for
"bucket" and you should see that discussion.

Is it *really* better for your users to see a low-relevance query
that happens to have the exact words in it before a very-high
ranking but not quite exact response?

Nopes. Thats the last thing my product manager will want.

Lets take an example to simplify this:

I have fields like title, description, tags. Now when I search for a term
"Indoor Photography" then I would like the results with exact match in
title to be
more important than in description or tags. However, if there is an exact
match in description
then it should be given more preference than the partial match in title.

Going by the points mentioned below and as per one of your posts
(
http://mail-archives.apache.org/mod_mbox/lucene-java-user/200609.mbox/%[email protected]%3E
)
I understand that I need to specify query time boosting like this:

title:Indoor Photography^2.5 description:Indoor Photography^1.5 tags:
Indoor Photography^1.2

That would go some distance towards what you want, but watch the syntax.
You might be better off constructing your own BooleanQuery. The syntax above
would actually parse something like title:Indoor default_field:Photography^
2.5. You
need parentheses. Also think about phrase queries....

Hope this helps
Erick

Let me know if this would help my cause.

Thnx for ur time n the valuable info.

--Rakesh S

Date: Fri, 21 Dec 2007 09:53:02 -0500
From: [email protected]
To: [email protected]
Subject: Re: Boosting Vs Sorting

OK, I'm trying to adjust to a Mac and my keyboard shortcuts sometimes
lead me to send the mail when I didn't intend. Sorry about that...

So, leaving aside how you form your "similar" query, I *think* you
want to form two clauses, your "exact" and your "similar" and
boost them individually, combined in a boolean query.

This will still interleave the results I think. But it's also a valid
question whether this is good or bad. Is it *really* better for your
users to see a low-relevance query that happens to have the exact
words in it before a very-high ranking but not quite exact response?
That, of course it up to your product manager....

If it is really a requirement, it seems to me that you would be able to
just form the two queries independently, then just post-process them.
One query is the exact version, and the second query is the similar one.
Then just combine the results as you please by iterating the hits
object for the exact query then following it by the same for the similar.
I don't see how sorting relates to your problem at all....

Best
Erick

On Dec 20, 2007 10:50 PM, Rakesh Shete wrote:

Hi all,

I am using Hibernate Search (http://www.hibernate.org/410.html)

which is

a wrapper around Lucene for performing search over info stored in

the DB. I

have questions related to Lucene boosting Vs sorting:

Is index time boosting of documents and fields better than

specifying

sorting parameters at search time?

I have been browsing through the Lucene mail archives for an answer

this. Going through them and reading on stuff related to Lucene

scoring, my

understanding is that if I know upfront at index time that the

relevance

order of results is based on certain fields, then, it is better to

have

index time boosting of documents and fields. Am I right here?

My requirements are like:
Results having an exact match to the input query string should have
highest preference followed by an exact match with field1, field2,

field3

and then followed by search query substring (or near match) match

with

field1, field2, field3.

Any suggestions are most welcome.

--Rakesh S

_________________________________________________________________
Post free property ads on Yello Classifieds now! www.yello.in
http://ss1.richmedia.in/recurl.asp?pid=219

_________________________________________________________________
Post free property ads on Yello Classifieds now! www.yello.in
http://ss1.richmedia.in/recurl.asp?pid=219

reply | permalink

Rakesh Shete

at Dec 24, 2007 at 1:29 pm

⇧

Hi Eric,

Thanks for the help. Most of it makes sense to me and I amgoing ahead with query boosting during search instead of index time boosting.

Could just tell me how do we arrive at the correct boost values for a query with multiple fields? My understanding is that this will vary as per the data that is being searched.

My query after boosting would look something like this:

+(i_title:sorting*^6.0 i_description:sorting*^3.5 i_detailedInfo:sorting*^1.5 i_tags:sorting*^1.2) -i_published:false +i_topicsClasses.id:1*

The reason for using high values is that I want to make sure that multiple occurrences of the search string in say field 'i_description' should not be treated as more relevant than a single match in the 'i_title' field.

Do you have some guideline using which I can arrive at the correct boost value in the above scenario?

--Rakesh S

Date: Fri, 21 Dec 2007 17:15:29 -0500
From: [email protected]
To: [email protected]
Subject: Re: Boosting Vs Sorting

See below...

On Dec 21, 2007 12:50 PM, Rakesh Shete wrote:

Hi Eric,

I don't see how sorting relates to your problem at all....

Is it *really* better for your users to see a low-relevance query
that happens to have the exact words in it before a very-high
ranking but not quite exact response?

Nopes. Thats the last thing my product manager will want.

Lets take an example to simplify this:

I have fields like title, description, tags. Now when I search for a term
"Indoor Photography" then I would like the results with exact match in
title to be
more important than in description or tags. However, if there is an exact
match in description
then it should be given more preference than the partial match in title.

Going by the points mentioned below and as per one of your posts
(
http://mail-archives.apache.org/mod_mbox/lucene-java-user/200609.mbox/%[email protected]%3E
)
I understand that I need to specify query time boosting like this:

title:Indoor Photography^2.5 description:Indoor Photography^1.5 tags:
Indoor Photography^1.2

Let me know if this would help my cause.

Thnx for ur time n the valuable info.

--Rakesh S

Date: Fri, 21 Dec 2007 09:53:02 -0500
From: [email protected]
To: [email protected]
Subject: Re: Boosting Vs Sorting

OK, I'm trying to adjust to a Mac and my keyboard shortcuts sometimes
lead me to send the mail when I didn't intend. Sorry about that...

So, leaving aside how you form your "similar" query, I *think* you
want to form two clauses, your "exact" and your "similar" and
boost them individually, combined in a boolean query.

This will still interleave the results I think. But it's also a valid
question whether this is good or bad. Is it *really* better for your
users to see a low-relevance query that happens to have the exact
words in it before a very-high ranking but not quite exact response?
That, of course it up to your product manager....

If it is really a requirement, it seems to me that you would be able to
just form the two queries independently, then just post-process them.
One query is the exact version, and the second query is the similar one.
Then just combine the results as you please by iterating the hits
object for the exact query then following it by the same for the similar.
I don't see how sorting relates to your problem at all....

Best
Erick

wrote:

Hi all,

I am using Hibernate Search (http://www.hibernate.org/410.html)

which is

a wrapper around Lucene for performing search over info stored in

the DB. I

have questions related to Lucene boosting Vs sorting:

Is index time boosting of documents and fields better than

specifying

sorting parameters at search time?

I have been browsing through the Lucene mail archives for an answer

this. Going through them and reading on stuff related to Lucene

scoring, my

understanding is that if I know upfront at index time that the

relevance

order of results is based on certain fields, then, it is better to

have

field3

and then followed by search query substring (or near match) match

with

_________________________________________________________________
Post free property ads on Yello Classifieds now! www.yello.in
http://ss1.richmedia.in/recurl.asp?pid=219

reply | permalink

Erick Erickson

at Dec 24, 2007 at 3:10 pm

⇧

I don't think there's any real formula for the "correct" boosts, it's more a
matter of
experimentation.

I'd look at some of the raw scores that come back (not the normalized ones)
and
go from there. See HitCollector (?).

Best
Erick

On Dec 24, 2007 8:29 AM, Rakesh Shete wrote:

Hi Eric,

Thanks for the help. Most of it makes sense to me and I amgoing ahead with
query boosting during search instead of index time boosting.

Could just tell me how do we arrive at the correct boost values for a
query with multiple fields? My understanding is that this will vary as per
the data that is being searched.

My query after boosting would look something like this:

+(i_title:sorting*^6.0 i_description:sorting*^3.5 i_detailedInfo:sorting*^
1.5 i_tags:sorting*^1.2) -i_published:false +i_topicsClasses.id:1*

The reason for using high values is that I want to make sure that multiple
occurrences of the search string in say field 'i_description' should not be
treated as more relevant than a single match in the 'i_title' field.

Do you have some guideline using which I can arrive at the correct boost
value in the above scenario?

--Rakesh S

Date: Fri, 21 Dec 2007 17:15:29 -0500
From: [email protected]
To: [email protected]
Subject: Re: Boosting Vs Sorting

See below...

On Dec 21, 2007 12:50 PM, Rakesh Shete wrote:

Hi Eric,

I don't see how sorting relates to your problem at all....

Could you just explain how is sorting different from boosting?

I have been trying to figure this out. Going through "Lucene In

Action" my

understanding of sorting is that it will kind of second level of

ordering

after the query results have been scored (Not sure if the relevance
established
by scoring is lost in this process).

Is it *really* better for your users to see a low-relevance query
that happens to have the exact words in it before a very-high
ranking but not quite exact response?

Nopes. Thats the last thing my product manager will want.

Lets take an example to simplify this:

I have fields like title, description, tags. Now when I search for a

term

"Indoor Photography" then I would like the results with exact match in
title to be
more important than in description or tags. However, if there is an

exact

match in description
then it should be given more preference than the partial match in

title.

Going by the points mentioned below and as per one of your posts
(

http://mail-archives.apache.org/mod_mbox/lucene-java-user/200609.mbox/%[email protected]%3E

)
I understand that I need to specify query time boosting like this:

title:Indoor Photography^2.5 description:Indoor Photography^1.5 tags:
Indoor Photography^1.2

That would go some distance towards what you want, but watch the syntax.
You might be better off constructing your own BooleanQuery. The syntax above
would actually parse something like title:Indoor
default_field:Photography^
2.5. You
need parentheses. Also think about phrase queries....

Hope this helps
Erick

Let me know if this would help my cause.

Thnx for ur time n the valuable info.

--Rakesh S

Date: Fri, 21 Dec 2007 09:53:02 -0500
From: [email protected]
To: [email protected]
Subject: Re: Boosting Vs Sorting

OK, I'm trying to adjust to a Mac and my keyboard shortcuts

sometimes

lead me to send the mail when I didn't intend. Sorry about that...

So, leaving aside how you form your "similar" query, I *think* you
want to form two clauses, your "exact" and your "similar" and
boost them individually, combined in a boolean query.

This will still interleave the results I think. But it's also a

valid

question whether this is good or bad. Is it *really* better for your
users to see a low-relevance query that happens to have the exact
words in it before a very-high ranking but not quite exact response?
That, of course it up to your product manager....

If it is really a requirement, it seems to me that you would be able

just form the two queries independently, then just post-process

them.

One query is the exact version, and the second query is the similar

one.

Then just combine the results as you please by iterating the hits
object for the exact query then following it by the same for the similar.
I don't see how sorting relates to your problem at all....

Best
Erick

since

you really want to compare field x across all documents. It seems
that query-time boosting is more relevant.

So, leaving aside how you form your "similar" q

On Dec 20, 2007 10:50 PM, Rakesh Shete < [email protected]>

wrote:

Hi all,

I am using Hibernate Search (http://www.hibernate.org/410.html)

which is

a wrapper around Lucene for performing search over info stored

the DB. I

have questions related to Lucene boosting Vs sorting:

Is index time boosting of documents and fields better than

specifying

sorting parameters at search time?

I have been browsing through the Lucene mail archives for an

answer

this. Going through them and reading on stuff related to Lucene

scoring, my

understanding is that if I know upfront at index time that the

relevance

order of results is based on certain fields, then, it is better

have

index time boosting of documents and fields. Am I right here?

My requirements are like:
Results having an exact match to the input query string should

have

highest preference followed by an exact match with field1,

field2,

field3

and then followed by search query substring (or near match)

match

with

field1, field2, field3.

Any suggestions are most welcome.

--Rakesh S

_________________________________________________________________

Post free property ads on Yello Classifieds now! www.yello.in
http://ss1.richmedia.in/recurl.asp?pid=219

_________________________________________________________________
Post free property ads on Yello Classifieds now! www.yello.in
http://ss1.richmedia.in/recurl.asp?pid=219

reply | permalink

Related Discussions

Discussion Navigation

view	thread \| post

Discussion Overview

group	java-user @
categories	lucene
posted	Dec 21, '07 at 3:51a
active	Dec 24, '07 at 3:10p
posts	7
users	2
website	lucene.apache.org

responses	expanded	Hotkey: s
font	variable	Hotkey: f
user style	avatars	Hotkey: a

Boosting Vs Sorting

Search Discussions

Related Discussions

2 users in discussion