Uploaded image for project: 'Lucene - Core'
  1. Lucene - Core
  2. LUCENE-3029

MultiPhraseQuery assigns different scores to identical docs when using 0 pos-incr


    • Bug
    • Status: Closed
    • Minor
    • Resolution: Fixed
    • None
    • 3.0.4, 3.2, 4.0-ALPHA
    • None
    • None
    • New


      If you have two identical docs with tokens a b c all zero pos-incr (ie
      they occur on the same position), and you run a MultiPhraseQuery with
      [a, b] and [c] (all pos incr 0)... then the two docs will get
      different scores despite being identical.

      Admittedly it's a strange query... but I think the scorer ought to
      count the phrase as having tf=1 for each doc.

      The problem is that we are missing a tie-breaker for the PhraseQuery
      used by ExactPhraseScorer, and so the PQ ends up flip/flopping such
      that every other document gets the same score. Ie, even docIDs all
      get one score and odd docIDs all get another score.

      Once I added the hard tie-breaker (ord) the scores are the same.

      However... there's a separate bug, that can over-count the tf, such
      that if I create the MPQ like this:

        mpq.add(new Term[] {new Term("field", "a")}, 0);
        mpq.add(new Term[] {new Term("field", "b"), new Term("field", "c")}, 0);

      I get tf=2 per doc, but if I create it like this:

        mpq.add(new Term[] {new Term("field", "b"), new Term("field", "c")}, 0);
        mpq.add(new Term[] {new Term("field", "a")}, 0);

      I get tf=1 (which I think is correct?).

      This happens because MultipleTermPositions freely returns the same
      position more than once: it just unions the positions of the two
      streams, so when both have their term at pos=0, you'll get pos=0
      twice, which is not good and leads to over-counting tf.

      Unfortunately, I don't see a performant way to fix that... and I'm not
      sure that it really matters that much in practice.


        1. LUCENE-3029.patch
          10 kB
          Michael McCandless



            mikemccand Michael McCandless
            mikemccand Michael McCandless
            0 Vote for this issue
            0 Start watching this issue