[
https://issues.apache.org/jira/browse/OPENNLP-1975?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18121747#comment-18121747
]
ASF GitHub Bot commented on OPENNLP-1975:
-----------------------------------------
krickert commented on PR #1314:
URL: https://github.com/apache/opennlp/pull/1314#issuecomment-5941893609
Looked at eval 93.
The Swedish quasi-Newton drop comes from this PR, and it is the cp change:
the default POS feature set includes `TokenClassFeatureGenerator`, and the
Swedish talbanken05 training data has exactly one token of the new kind, `Å.`
(the old set had Ä Ö Ü but not Å). The test set has none. With that one feature
different, the quasi-Newton trainer converges to a slightly different model,
0.93405 instead of 0.93777.
Three checks say it is the optimizer path and the feature carries the same
information as before: the Swedish GIS model trained on the same features keeps
its pinned accuracy; main at this PR's base and main before #1321 both
reproduce the old value locally; and this head under
`-Dopennlp.compat.mode=LEGACY` returns the old value to the last digit.
So I re-pinned the QN expectation to the new value, with that reasoning in
the commit. The eval needs one more run on the new head.
> RegEx removal part 9: Correct the output of the regex sites in
> TokenSampleStream, ParserTool, ConlluStream, DownloadUtil and
> FeatureGeneratorUtil
> -------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: OPENNLP-1975
> URL: https://issues.apache.org/jira/browse/OPENNLP-1975
> Project: OpenNLP
> Issue Type: Bug
> Reporter: Richard Zowalla
> Assignee: Kristian Rickert
> Priority: Major
> Fix For: 3.0.0
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> Split out of OPENNLP\-1928 \(PR #1275\). Depends on OPENNLP\-1928 \(RegEx
> removal part 1\).
> Replaces the regular expressions of these sites with scans that follow the
> intended format instead of copying the regex quirks. The output changes:*
> {{TokenSampleStream}} \(English\): tokens are split on Unicode White\_Space;
> a leading space no longer yields an empty first token; a line without a token
> resets the open\-quote state; a token with a letter or digit of any script is
> a word.
> * {{ParserTool}}: brackets that follow each other are each separated
> \({{a\)\)b}} gives {{a \) \) b}}\).
> * {{ConlluStream}}: a malformed multiword token id \({{3\-}}, {{\-4}},
> {{3\-\-4}}, {{1\-2\-3}}, {{4\-3}}\) fails with an {{InvalidFormatException}}
> naming the id.
> * {{DownloadUtil}}: a blank checksum file is an {{IOException}} saying so.
> * {{FeatureGeneratorUtil}}: the {{cp}} class requires exactly one capital and
> one period.
> Before merge: tokenizer eval build that reads through {{TokenSampleStream}}.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
