lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Chris Hostetter <hossman_luc...@fucit.org>
Subject Re: faceting and categorizing on color?
Date Wed, 21 Jun 2006 08:24:46 GMT

: I thought that having: F00000 FF0000 FFFFFF FFFF00...
: in one field and then searching for FFFFFF in it would
: match all documents that contain that "word" so I
	...
: the counts were equal. I guess I am still not clear on
: what the differences/advantages/disadvantages are
: between the above an document and one that like this:
	...
: color in its own field? So perhaps a more general
: question is when is it better to collapse a bunch of
: words into a single field vs. spread them out over
: many fields, all of which have the same field name?

My point was that (depending on what exactly you were trying to express)
when you try to represent a document's structure in email like that both
examples have the exact same structure in the index.  For any given field
name, there is only one Field in the index, and it has a stream of values
regardless of how you constructed your document object -- technically
there isn't even really a "field" as much as there is an ordered list of
Terms.  Specificly, this...

  IndexWriter w = new IndexWriter(d, new WhitespaceAnalyzer(), true);
  Document doc = new Document();
  doc.add(new Field("color", "red blue green", Store.NO, Index.TOKENIZED));
  w.addDocument(doc);

...is going to produce and index with the exact same terms maped to that
document as this...

  IndexWriter w = new IndexWriter(d, new WhitespaceAnalyzer(), true);
  Document doc = new Document();
  doc.add(new Field("color", "red", Store.NO, Index.TOKENIZED));
  doc.add(new Field("color", "blue", Store.NO, Index.TOKENIZED));
  doc.add(new Field("color", "green", Store.NO, Index.TOKENIZED));
  w.addDocument(doc);

...the only differnece that might pop up, is if you configure your
analyzer to have a positionIncrimentGap greater then 0, in which case
there will be a bigger "gap" between the red and blue and green in the
second case then in the fist (which will affect any sloppy searches that
you might do)

LIA has a lot of gret info on exactly how thisworks, and can relaly help
you "think" in terms of Terms.

: That's a great idea and seems way more straightforward
: than the RGB/HSV etc. distance calculation algorithms
: I've been reading about :o) I'll have to run some
: tests to see how accurate that reduction appears to
: people.

that's just an easy one if you are already representing your colors in RGB
hex codes -- the bottom line is any method you have for simplifying your
pallet will make it easier to do facet counts (because you ar reducing the
number of facets)

: Huh, perhaps I don't understand the HitCollector
: fully. Are you saying that if I have an index with 100
: documents, each of which have a color field (let's say
: 25 of the documents have FFFFFF in the color field)
: and I do a search for FFFFFF...using a HitCollector
: I'm iterating over all 100 documents while extracting
: values whereas with the regular Hits based search I
: would only be iterating over 25?

First off, forget the Hits class exists -- it has no place in a discussion
of facet counts where you need to look at every document that matches a
given query.

Second: my point is that when you are starting with an arbitrary query,
and then you wnat to provide counts for a bunch of facet values (ie:
colors) you have no idea which facet values are the "best" -- you have no
way of knowing that you should start with FFFFFF, you have to try them
all...

Conceptually facets are about set intersections and the cardinality of
those intersections.  You have a main result set of documents which match
your user's query, and you have many hypothetical sets of documents that
each represent all docs that contain some single color that deifines that
set.  You want to intersect each of those sets with your main result set,
and find out which colors product the greatest intersection cardinality.

With the BitSet approach, you can directly implimenting this comceptual
model -- making real sets out of your hypothetical sets.  But if that's
not feasible from a memory usage perspective, then you can traverse the
two dimensions of the problem space (colors and documents) in either
order:
 1) you can use the FieldCache (or some other data structure you build
from TermEnums) to have fast access to the list of colors in each doc, and
then in your HitCollector as you collect each docId matching your primary
query, incriment a counter for the corrisponding colors.
 2) you can start with a HitCollector that just records all of the docIds
that match your primary query -- a BitSet is convinient for this -- and
then use a TermEnum with a TermDocs to iterate over every color in your
index, and count up how many of the documents those colors map to are in
your BitSet

(NOTE: using a FieldCache for this is non-trivial since FieldCache only
works if each doc has at most one value, i wasn't really thinking about
the specifics of your color example when i mentioned it before -- but the
concept is still the same: an array that allows fast lookup by docId, you
could make your own array containing the arrays of values just as easily
as the FieldCache)





-Hoss


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