[
https://issues.apache.org/jira/browse/LUCENE-8985?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17189288#comment-17189288
]
Michael McCandless commented on LUCENE-8985:
--------------------------------------------
Thank you [~nppoly] for the thorough PR – I will try to review again soon.
Holes are a challenge for graph token filters. I have long felt that stop
words (and other holey tokens) should not be removed from the token stream, but
rather a new {{DeletedAttribute}} would mark the token as deleted (and not to
be indexed) but still the token would remain (and be {{incrementToken}}'d) to
record the metadata about that token for the sake of future token filter
stages. E.g. this would allow analyzers that mark stopwords for deletion but
then e.g. a {{SynonymGraphFilter}} could still apply over the stopwords, e.g.
{{lord of the rings}} could still match properly even if {{of}} and {{the}}
were marked as deleted. But, that is a much larger change, and no need to hold
up this first approach for that!
> SynonymGraphFilter cannot handle input stream with tokens filtered.
> -------------------------------------------------------------------
>
> Key: LUCENE-8985
> URL: https://issues.apache.org/jira/browse/LUCENE-8985
> Project: Lucene - Core
> Issue Type: Bug
> Reporter: Chongchen Chen
> Priority: Major
> Fix For: 8.3
>
> Attachments: SGF_SF_interaction.patch.txt
>
> Time Spent: 4h
> Remaining Estimate: 0h
>
> [~janhoy] find the bug.
> In an analyzer with e.g. stopFilter where tokens are removed from the stream
> and replaced with a “hole”, synonymgraphfilter will not preserve these holes
> but remove them, resulting in certain phrase queries failing.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]