lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Erick Erickson <erickerick...@gmail.com>
Subject Re: Getting a score of a specific document
Date Mon, 18 May 2009 13:07:04 GMT
As best I understand it, you DO NOT WANT A FILTER. Filters do notcontribute
to scoring, therefore do not rank your documents. If you use
a filter, the most irrelevant document could be first. You want to use
a HitCollector, see the link in my last e-mail. That link includes an
example of using a bitset, which you can create pretty easily from your
list of document IDs.

Best
Erick

On Mon, May 18, 2009 at 2:55 AM, liat oren <oren.liat@gmail.com> wrote:

> Sorry I didn't explain myself well.
>
> The problem I try to address is the following:
> Think about the case where you have 100,000 documents indexed. Take word
> 'a'
> - if it appears in 80,000 documents, you want the score to take it into
> account. You want only to see how 20,000 documents are close to a query,
> and
> only 10,000 of these contain the word 'a'.
> 80,000 / 100,000 (the 'statistics' of the whole index) is much smaller than
> 10,000 / 20,000 (the 'statistics of only the group of documents). So it
> does
> affect the score if I use the whole index or just the documents I am
> interested in.
> It might be that the order of these desired documents will not change, but
> I
> don;t see how you can assure it since the idf value can be really
> different.
>
> So, I want the documents for my query to be
> ranked *relative to each other*, AND NOT restricted to only the documents
> I care about.
> For that case, I need to use the filter, right?
>
> Its fine if I get the results in DocumentID - then I open these using
> IndexReader to get the fields I need.
>
> Could you please give me an example of how I creat the Filter that filters
> out a given list of ids?
>
> Thanks!
> Liat
> 2009/5/18 Erick Erickson <erickerickson@gmail.com>
>
> > I'm still unclear what you want the statistics *for*. "statistics"
> > are pretty meaningless as far as I understand. The whole point
> > of scoring is to use various "statistics" to *rank* documents *for
> > a specific query*. You cannot, for instance, compare scores
> > between different queries in any meaningful way.
> >
> > If you're saying that you want the documents for your query to be
> > ranked *relative to each other*, but restricted to only the documents
> > you care about, then I think you need a HitCollector
> > because a Filter (last I knew) doesn't score documents therefore
> > won't order them.
> >
> > But asking if the statistics reflect the whole index just isn't making
> any
> > sense to me. If you're asking that question I suspect that there's
> > something about your problem space I don't understand and
> > you're not explaining simply enough for me to grasp <G>.
> >
> > So forget a Filter because you'll get the documents back in
> > (probably, but my memory is weak some days) document ID
> > order. Implement a HitCollector whose collect method only
> > sets bits for docs in your list. See:
> >
> >
> http://lucene.apache.org/java/2_2_0/api/org/apache/lucene/search/HitCollector.html#collect(int,%20float)
> >
> > Best
> > Erick
> >
> >
> > On Sun, May 17, 2009 at 3:57 AM, liat oren <oren.liat@gmail.com> wrote:
> >
> > > Yes, this is what I need - I don't need to get the scores for the
> > documents
> > > that were filtered.
> > > The statistics I ment are idf(t) for example.
> > > I want these to include the whole index of course.
> > > It will include this info of all the index, right?
> > >
> > > if I have a list of ids that the query should look at, which Filter
> > should
> > > I
> > > use?
> > >
> > > Thanks a lot,
> > > Liat
> > >
> > > 2009/5/14 Erick Erickson <erickerickson@gmail.com>
> > >
> > > > Hmmm, come to think of it, if you pass the Filter to the search
> > I*think*
> > > > you
> > > > don't get scores for that clause, but you may want to
> > > > check it out...
> > > >
> > > > So I think you should think about implementing a HitCollector
> > > > and collect only the documents you care about.
> > > >
> > > > This is really very little extra work since all the documents have
> > > > to be evaluated anyway.
> > > >
> > > > I'm not sure what you mean by statistics for the whole index. I
> suspect
> > > > you're wondering if the scores reflect all the documents. But you
> don't
> > > > care because scores are not relevant between different queries, and
> > > > if they are calculated only within the query you're running, all the
> > > > documents returned have scores that rank them relative to each other.
> > > >
> > > > Best
> > > > Erick
> > > >
> > > > On Thu, May 14, 2009 at 9:16 AM, liat oren <oren.liat@gmail.com>
> > wrote:
> > > >
> > > > > Yes, I have a pre-defined list of documents that I care about.
> > > > > Then I can do the search on these, but it will take the statictics
> of
> > > the
> > > > > whole index, right?
> > > > >
> > > > >
> > > > >
> > > > >
> > > > > 2009/5/14 Erick Erickson <erickerickson@gmail.com>
> > > > >
> > > > > > I don't know if I'm understanding what you want, but if you
havea
> > > > > > pre-defined list of documents, couldn't you form a Filter? Then
> > > > > > your results would only be the documents you care about.
> > > > > >
> > > > > > If this is irrelevant, perhaps you could explain a bit more
about
> > > > > > the problem you're trying to solve.
> > > > > >
> > > > > > Best
> > > > > > Erick
> > > > > >
> > > > > > On Thu, May 14, 2009 at 5:03 AM, liat oren <oren.liat@gmail.com>
> > > > wrote:
> > > > > >
> > > > > > > Hi,
> > > > > > >
> > > > > > > I have a big index and I want to get for a specific search
only
> > the
> > > > > > grades
> > > > > > > of a list of documents.
> > > > > > > Is there a better way to get this score than looping on
all the
> > > > > reasults
> > > > > > > set?
> > > > > > >
> > > > > > > Thanks,
> > > > > > > Liat
> > > > > > >
> > > > > >
> > > > >
> > > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message