[
https://issues.apache.org/jira/browse/LUCENE-6717?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14660104#comment-14660104
]
Adrien Grand commented on LUCENE-6717:
--------------------------------------
I like it too, I'm curious how it compared to PhraseQuery now. :-)
I think we should set cost=required.cost() when there are required terms,
otherwise this query will still return the same cost as a disjunction even
though it can do much better.
I'm curious about the assertion at the beginning of the doNext() method, it's
been both changed and commented out, should we just remove it if invariants are
hard to verify?
TermAutomatonQuery.termIsRequired is documented as public for testing, but it
looks to me that pkg-private would be enough?
> TermAutomatonQuery should be two-phased
> ---------------------------------------
>
> Key: LUCENE-6717
> URL: https://issues.apache.org/jira/browse/LUCENE-6717
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Michael McCandless
> Assignee: Michael McCandless
> Attachments: LUCENE-6717.patch
>
>
> {{TermAutomatonQuery}} (still in sandbox) is a simple way to get accurate
> query-time multi-token synonyms using the new {{SynonymGraphFilter}} from
> LUCENE-6664. It already has a utility class to directly translate an
> incoming {{TokenStream}} into a corresponding query.
> However the query is likely quite slow because it always iterates positions
> for all terms in the automaton.
> I think one simple approach is to walk the automaton and find the subset of
> terms (if any) that appear in common to all paths, and then approximate with
> {{ConjunctionDISI}} like {{PhraseQuery}} does. Such a subset doesn't always
> exist for an automaton (i.e. it could be empty), so the logic would have to
> be conditional...
> And I think there are more complex approximations we could make, but using
> {{ConjunctionDISI}} seems like a simple start.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
