lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Uwe Schindler (JIRA)" <>
Subject [jira] [Commented] (LUCENE-4766) Pattern token filter which emits a token for every capturing group
Date Wed, 13 Feb 2013 13:32:13 GMT


Uwe Schindler commented on LUCENE-4766:

bq. Clinton, I think you can trash the offset attribute reference in there entirely just don't
mess with them at all.

That's part of a bigger problem in the current code. The idea of this filter is to make from
one input Token multiple output Tokens. To make this work correct, the *new* output tokens
must be produced based on the original token (means the filter must reset the new produced
token to a clean state, otherwise it might happen that unrelated and unknown attributes stay
alive with wrong values - especiall if later TokenFilter change attributes, e.g. a Synonymfilter
is inserting more synonyms). The problem Clinton had was that he had to re-set the offset
attribute (although he does not change it); but he missed possible other attributes on the
stream he does not know about.

If you look at other filters doing similar things like Synonymfilter, WDF, the way it has
to work is like that:
- The first token emmitted is the "original one, maybe modified
- All "inserted Tokens" are cloned from the original (first) token, use captureState/restoreState
to do that. This will initialize the attribute source to the exact same token like the original
(unmodified one). After you called restoreState, you can *modify* the attribute (like term
text) and setPositionIncrement(0). You can then leave the the offset (and other unknown attributes
that may be on the token stream) unchanged - don't reference them at all.
> Pattern token filter which emits a token for every capturing group
> ------------------------------------------------------------------
>                 Key: LUCENE-4766
>                 URL:
>             Project: Lucene - Core
>          Issue Type: New Feature
>          Components: modules/analysis
>    Affects Versions: 4.1
>            Reporter: Clinton Gormley
>            Assignee: Simon Willnauer
>            Priority: Minor
>              Labels: analysis, feature, lucene
>             Fix For: 4.2
>         Attachments: LUCENE-4766.patch, LUCENE-4766.patch, LUCENE-4766.patch, LUCENE-4766.patch
> The PatternTokenizer either functions by splitting on matches, or allows you to specify
a single capture group.  This is insufficient for my needs. Quite often I want to capture
multiple overlapping tokens in the same position.
> I've written a pattern token filter which accepts multiple patterns and emits tokens
for every capturing group that is matched in any pattern.
> Patterns are not anchored to the beginning and end of the string, so each pattern can
produce multiple matches.
> For instance a pattern like :
> {code}
>     "(([a-z]+)(\d*))"
> {code}
> when matched against: 
> {code}
>     "abc123def456"
> {code}
> would produce the tokens:
> {code}
>     abc123, abc, 123, def456, def, 456
> {code}
> Multiple patterns can be applied, eg these patterns could be used for camelCase analysis:
> {code}
>     "([A-Z]{2,})",
>     "(?<![A-Z])([A-Z][a-z]+)",
>     "(?:^|\\b|(?<=[0-9_])|(?<=[A-Z]{2}))([a-z]+)",
>     "([0-9]+)"
> {code}
> When matched against the string "letsPartyLIKEits1999_dude", they would produce the tokens:
> {code}
>     lets, Party, LIKE, its, 1999, dude
> {code}
> If no token is emitted, the original token is preserved. 
> If the preserveOriginal flag is true, it will output the full original token (ie "letsPartyLIKEits1999_dude")
in addition to any matching tokens (but in this case, if a matching token is identical to
the original, it will only emit one copy of the full token).
> Multiple patterns are required to allow overlapping captures, but also means that patterns
are less dense and easier to understand.
> This is my first Java code, so apologies if I'm doing something stupid.

This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see:

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message