[
https://issues.apache.org/jira/browse/OPENNLP-338?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13438538#comment-13438538
]
Joern Kottmann commented on OPENNLP-338:
----------------------------------------
The best F-Measure on the conll data i could get was 0.05.
Here are the details:
Precision: 0.38461538461538464
Recall: 0.02844950213371266
F-Measure: 0.05298013245033113
In the QNModel the case that a feature does not has a param is much more
frequent in the name finder than in the PPA data set I think. Another
difference is that it uses beam search to look for a likely sequence. Maybe the
SMOOTHING_VALUE needs to be adjusted. I changed it to 1 and used a cutoff of 5
instead for the test results above.
Is the smoothing needed? Should it be disabled?
Do we need to initialize the results array in QNModel.eval with some prior
values, like we in the GISModel?
Should the CONVERGE_TOLERANCE in QNTrainer be lower?
> Add L-BFGS parameter estimation training to maxent
> --------------------------------------------------
>
> Key: OPENNLP-338
> URL: https://issues.apache.org/jira/browse/OPENNLP-338
> Project: OpenNLP
> Issue Type: New Feature
> Components: Maxent
> Reporter: Joern Kottmann
> Assignee: Hyosup Shim
> Fix For: tools-1.5.3
>
> Attachments: nl-per.testa, nl-per.testb, nl-per.train,
> patch20120814-lbfgs.txt, patch20120821-lbfgs, patch-lbfgs.txt
>
>
> Add support for the L-BFGS algorithm to train a maxent classifier.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
