[
https://issues.apache.org/jira/browse/LUCENE-9112?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17025816#comment-17025816
]
Markus Jelsma commented on LUCENE-9112:
---------------------------------------
I discovered a problem with my sentence detector model and retrained to correct
for it. The model now properly splits sentences where they should when running
it in OpenNLP.
The problem with SegmentingTokenizerBase persists with the correct models.
> SegmentingTokenizerBase splits terms that occupy 1024th positions in text
> -------------------------------------------------------------------------
>
> Key: LUCENE-9112
> URL: https://issues.apache.org/jira/browse/LUCENE-9112
> Project: Lucene - Core
> Issue Type: Bug
> Components: modules/analysis
> Affects Versions: master (9.0)
> Reporter: Markus Jelsma
> Priority: Major
> Labels: opennlp
> Fix For: master (9.0)
>
> Attachments: LUCENE-9112-unittest.patch, LUCENE-9112-unittest.patch,
> en-sent.bin, en-token.bin
>
>
> The OpenNLP tokenizer show weird behaviour when text contains spurious
> punctuation such as having triple dots trailing a sentence...
> # the first dot becomes part of the token, having 'sentence.' becomes the
> token
> # much further down the text, a seemingly unrelated token is then suddenly
> split up, in my example (see attached unit test) the name 'Baron' is split
> into 'Baro' and 'n', this is the real problem
> The problems never seem to occur when using small texts in unit tests but it
> certainly does in real world examples. Depending on how many 'spurious' dots,
> a completely different term can become split, or the same term in just a
> different location.
> I am not too sure if this is actually a problem in the Lucene code, but it is
> a problem and i have a Lucene unit test proving the problem.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
