I wanted to check my understanding of phrase queries vs. span queries.
My problem is that given the simple terms the user enters, I want to
check for phrases made from those terms with some slop, BUT MAINTAIN THE
ORDER OF THE TERMS.
Is my understanding correct, I can NOT use match_phrase, because once I
start playing with a slop, I may end up scoring an out of order phrase.
Is the only choice in my case a span_near?
Another question: Are my chains of span_nears (see below) the only way
to get this kind of scoring?
I want what can be summarized as
terms are good, in fact they are required;
a 2-word phrase is better;
longer phrases even better;
Just all the terms only beat even a short 2-word phrase when there is
really a lot of them. Short phrases beat longer phrases similarly -
only when there are really a lot of them.
If the words are: aaa bbb ccc
I might generate a LUCENE query like
+aaa +bbb +ccc "aaa bbb"^100~6 "aaa bbb ccc"^10000~6
"aaa 1 2 3 4 5 6 7 8 9 10 11 ... 110 111 112 bbb" is a matching
document, but
"aaa 1 2 3 4 5 bbb" is better, but
"aaa 1 2 3 bbb 1 2 3 4 5 ccc" is lots (i.e. 100x!) better (each term
within 6 of its neighbor).
When using Lucene (I thought) I found that phrase queries with slop do
not limit themselves to preserving the order.
The Lucene PhraseQuery spells this out in the discussion of slop.
https://lucene.apache.org/core/3_6_1/api/core/org/apache/lucene/search/PhraseQuery.html#setSlop(int)
So I switched to SpanNearQuery, one for each phrase - a 2-term
span_near, a 3-term span_near, etc.
It has all of what I need preserve order, specify a slop, and a boost.
When I ported to ES I used the same span near query.
http://www.elasticsearch.org/guide/reference/query-dsl/span-near-query.html
Am I missing something, can I use match_phrase or some other query in ES
to get what I want?
Is my manual application of phrase boost really necessary?
From
http://www.elasticsearch.org/guide/reference/query-dsl/match-query.html
I see slop and order are only addressed by the sentence
"A phrase query maintains order of the terms up to a configurable |slop|
(which defaults to 0)."
That sounds like it will allow out of order which matches my
understanding of Lucene PhraseQuery.
Suggestions or comments about the use of chains of span_nears would be
appreciated.
-Paul
--