[
https://issues.apache.org/jira/browse/SOLR-16594?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17739237#comment-17739237
]
Gus Heck commented on SOLR-16594:
---------------------------------
I fiddled around with customizing solr to do this and increasingly felt that
the real fundamental problem was that lucene was throwing out the offset
information that it certainly has available at analysis time and that it would
be better if lucene just kept that info around for the parser. So I spent some
time and came up with a patch to lucene (vs 9.5 since that's what current solr
was depending on when I started) with the following concepts/goals
BEFORE looking at the link below please read the following list which should
clarify some oddities you may note:
# Only touch the query world, minimize or hopefully avoid changes relating to
the indexing side of lucene.
# To facilitate #1 allow term to be subclassed (remove final from the class)
# Create a QueryTerm class that constructs with a value representing the start
offset. The end offset isn't needed here, and so that would be an exercise for
someone who finds a use for it.
# Offset information isn NOT meant to remain reliable after rewrite() is
called. Parser's shouldn't need this, and there were places where rewrite
related code seemed to touch indexing related classes that I wanted to avoid.
# All tests should pass with only API change related modification (i.e.
providing the offset to ensure equality.)
# Code not involved in parsing, or parsers not extending QueryBuilder that
don't care about term offset (i.e don't support min should match) can safely
just set the offset to zero and never worry about it. A convenience constructor
to do this will keep things backwards compatible
# Convenience constructors to automatically set offset to zero were explicitly
avoided on a temporary basis despite this creating a lot of work. This is to
force correct integration into solr (will cause compile errors that can be
easily spotted and resolved)
# After Solr integration is complete the convenience constructor (setting the
offset to zero) should be added and utilized through the code base (this will
shrink the diff significantly).
All tests pass.
Point 5 failed on one test (QueryDecomposerTest) which had several cases
comparing the output of parsing two entirely different queries, and thus the
offsets must (correctly) differ and equality was lost. For those cases I
constructed the expected case directly instead of parsing a string for it so
that the correct offsets could be supplied.
[Here|https://github.com/kmwtechnology/lucene/commit/8c3773b28acb8767fb4796fe21295be5d62d8936]
is the diff for a commit of this code prior to integration into solr (there
could be a change or two needed to ensure that is smooth, TBD)
Naming things is hard. I've named the key class
[QueryTerm|https://github.com/kmwtechnology/lucene/commit/8c3773b28acb8767fb4796fe21295be5d62d8936#diff-5ef3fbdabc4dff41320f238b58d2032a468b0cec897cf5057285b78b47348d2b]
right now, but I also Toyed with ParsedTerm (less confusion with TermQuery),
but that seemed less appropriate overall. Naming suggestions welcome of course.
I'll make a PR out of this shortly. Mostly want a chance for at least a few
folks to see the motivation before they stumble on the mail for the large and
hard to read PR and start forming opinions with half the information.
Also once Solr integration is worked out, I would fully expect to run lucene
benchmarks with/without to ensure this didn't have unexpected effects there. It
would be a Lucene 10 (and thus probably solr 10) solution I believe.
> improve eDismax strategy for generating a term-centric query
> ------------------------------------------------------------
>
> Key: SOLR-16594
> URL: https://issues.apache.org/jira/browse/SOLR-16594
> Project: Solr
> Issue Type: Improvement
> Components: query parsers
> Reporter: Rudi Seitz
> Assignee: Gus Heck
> Priority: Major
> Time Spent: 40m
> Remaining Estimate: 0h
>
> When parsing a multi-term query that spans multiple fields, edismax attempts
> to generate a term-centric query structure
>
> sometimes switches from a "term-centric" to a "field-centric" approach. This
> creates inconsistent semantics for the {{mm}} or "min should match" parameter
> and may have an impact on scoring. The goal of this ticket is to improve the
> approach that edismax uses for generating term-centric queries so that
> edismax would less frequently "give up" and resort to the field-centric
> approach. Specifically, we propose that edismax should create a dismax query
> for each distinct startOffset found among the tokens emitted by the field
> analyzers. Since the relevant code in edismax works with Query objects that
> contain Terms, and since Terms do not hold the startOffset of the Token from
> which Term was derived, some plumbing work would need to be done to make the
> startOffsets available to edismax.
>
> BACKGROUND:
>
> If a user searches for "foo bar" with {{{}qf=f1 f2{}}}, a field-centric
> interpretation of the query would contain a clause for each field:
> {{ (f1:foo f1:bar) (f2:foo f2:bar)}}
> while a term-centric interpretation would contain a clause for each term:
> {{ (f1:foo f2:foo) (f1:bar f2:bar)}}
> The challenge in generating a term-centric query is that we need to take the
> tokens that emerge from each field's analysis chain and group them according
> to the terms in the user's original query. However, the tokens that emerge
> from an analysis chain do not store a reference to their corresponding input
> terms. For example, if we pass "foo bar" through an ngram analyzer we would
> get a token stream containing "f", "fo", "foo", "b", "ba", "bar". While it
> may be obvious to a human that "f", "fo", and "foo" all come from the "foo"
> input term, and that "b", "ba", and "bar" come from the "bar" input term,
> there is not always an easy way for edismax to see this connection. When
> {{{}sow=true{}}}, edismax passes each whitespace-separated term through each
> analysis chain separately, and therefore edismax "knows" that the output
> tokens from any given analysis chain are all derived from the single input
> term that was passed into that chain. However, when {{{}sow=false{}}},
> edismax passes the entire multi-term query through each analysis chain as a
> whole, resulting in multiple output tokens that are not "connected" to their
> source term.
> Edismax still tries to generate a term-centric query when {{sow=false}} by
> first generating a boolean query for each field, and then checking whether
> all of these per-field queries have the same structure. The structure will
> generally be uniform if each analysis chain emits the same number of tokens
> for the given input. If one chain has a synonym filter and another doesn’t,
> this uniformity may depend on whether a synonym rule happened to match a term
> in the user's input.
> Assuming the per-field boolean queries _do_ have the same structure, edismax
> reorganizes them into a new boolean query. The new query contains a dismax
> for each clause position in the original queries. If the original queries are
> {{(f1:foo f1:bar)}} and {{(f2:foo f2:bar)}} we can see they have two clauses
> each, so we would get a dismax containing all the first position clauses
> {{(f1:foo f1:bar)}} and another dismax containing all the second position
> clauses {{{}(f2:foo f2:bar){}}}.
> We can see that edismax is using clause position as a heuristic to reorganize
> the per-field boolean queries into per-term ones, even though it doesn't know
> for sure which clauses inside those per-field boolean queries are related to
> which input terms. We propose that a better way of reorganizing the per-field
> boolean queries is to create a dismax for each distinct startOffset seen
> among the tokens in the token streams emitted by each field analyzer. The
> startOffset of a token (rather, a PackedTokenAttributeImpl) is "the position
> of the first character corresponding to this token in the source text".
> We propose that startOffset is a resonable way of matching output tokens up
> with the input terms that gave rise to them. For example, if we pass "foo
> bar" through an ngram analysis chain we see that the foo-related tokens all
> have startOffset=0 while the bar-related tokens all have startOffset=4.
> Likewise, tokens that are generated via synonym expansion have a startOffset
> that points to the beginning of the matching input term. For example, if the
> query "GB" generates "GB gib gigabyte gigabytes" via synonym expansion, all
> of those four tokens would have startOffset=0.
> Here's an example of how the proposed edismax logic would work. Let's say a
> user searches for "foo bar" across two fields, f1 and f2, where f1 uses a
> standard text analysis chain while f2 generates ngrams. We would get
> field-centric queries {{(f1:foo f1:bar)}} and ({{{}f2:f f2:fo f2:foo f2:b
> f2:ba f2:bar){}}}. Edismax's "all same query structure" check would fail
> here, but if we look for the unique startOffsets seen among all the tokens we
> would find offsets 0 and 4. We could then generate one clause for all the
> startOffset=0 tokens {{(f1:foo f2:f f2:fo f2:foo)}} and another for all the
> startOffset=4 tokens: {{{}(f1:bar f2:b f2:ba f2:bar){}}}. This would
> effectively give us a "term-centric" query with consistent mm and scoring
> semantics, even though the analysis chains are not "compatible."
> As mentioned, there would be significant plumbing needed to make startOffsets
> available to edismax in the code where the per-field queries are converted
> into per-term queries. Modifications would possibly be needed in both the
> Solr and Lucene repos. This ticket is logged in hopes of gathering feedback
> about whether this is a worthwhile/viable approach to pursue further.
>
> Related tickets:
> https://issues.apache.org/jira/browse/SOLR-12779
> https://issues.apache.org/jira/browse/SOLR-15407
>
> Related blog entries:
> [https://opensourceconnections.com/blog/2018/02/20/edismax-and-multiterm-synonyms-oddities]
> [https://sease.io/2021/05/apache-solr-sow-parameter-split-on-whitespace-and-multi-field-full-text-search.html]
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@solr.apache.org
For additional commands, e-mail: issues-h...@solr.apache.org
