[
https://issues.apache.org/jira/browse/OPENNLP-777?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14963353#comment-14963353
]
Cohan Sujay Carlos commented on OPENNLP-777:
--------------------------------------------
[~joern] and [~teofili],
Just a quick update, since it's been a month since the last message on this
thread.
I'm finding the setup required to replicate Joern's environment a real bear
(I'm sorry but I am just not familiar with the build tools), but I'll continue
banging at it next week.
But is there any way that either [~joern] or [~teofili] could check if the
Naive Bayes classifier is indeed being used when you call the model as Joern
has described using NaiveBayesTrainerParams.txt?
It just seems very unlikely that the NB model would - if it is being called -
fare worse than a perceptron at this task - given its superior performance in
the PrepAttach testcase.
Also since [~joern] is working on a language identifier, I thought I should
mention that sequential models (markov chains) with character level features
fare far better than linear classifiers with word-level features at that task
... a really good method is described by Sibun and Reynar in their paper
"Language Identification: Examining the Issues". The training command that
Joern used above doesn't seem to specify character-level features, so I am
assuming he's using word-level features and a linear classifier - and that
wouldn't work well in any case.
> Naive Bayesian Classifier
> -------------------------
>
> Key: OPENNLP-777
> URL: https://issues.apache.org/jira/browse/OPENNLP-777
> Project: OpenNLP
> Issue Type: New Feature
> Components: Machine Learning
> Environment: J2SE 1.5 and above
> Reporter: Cohan Sujay Carlos
> Assignee: Tommaso Teofili
> Priority: Minor
> Labels: NBClassifier, bayes, bayesian, classifier, multinomial,
> naive, patch
> Attachments: D1TopicClassifierTrainingDemoNB.java,
> D1TopicClassifierUsageDemoNB.java, NaiveBayesCorrectnessTest.java,
> naive-bayesian-classifier-for-opennlp-1.6.0-rc6-with-test-cases.patch,
> prep-attach-test-case-for-naive-bayesian-classifier-for-opennlp-1.6.0-rc6.patch,
> topics.train
>
> Original Estimate: 504h
> Remaining Estimate: 504h
>
> I thought it would be nice to have a Naive Bayesian classifier in OpenNLP (it
> lacks one at present).
> Implementation details: We have a production-hardened piece of Java code for
> a multinomial Naive Bayesian classifier (with default Laplace smoothing) that
> we'd like to contribute. The code is Java 1.5 compatible. I'd have to write
> an adapter to make the interface compatible with the ME classifier in
> OpenNLP. I expect the patch to be available 1 to 3 weeks from now.
> Below is the email trail of a discussion in the dev mailing list around this
> dated May 19th, 2015.
> <snip>
> Tommaso Teofili via opennlp.apache.org
> to dev
> Hi Cohan,
> I think that'd be a very valuable contribution, as NB is one of the
> foundation algorithms, often used as basis for comparisons.
> It would be good if you could create a Jira issue and provide more details
> about the implementation and, eventually, a patch.
> Thanks and regards,
> Tommaso
> </snip>
> 2015-05-19 9:57 GMT+02:00 Cohan Sujay Carlos
> > I have a question for the OpenNLP project team.
> >
> > I was wondering if there is a Naive Bayesian classifier implementation in
> > OpenNLP that I've not come across, or if there are plans to implement one.
> >
> > If it is the latter, I should love to contribute an implementation.
> >
> > There is an ME classifier already available in OpenNLP, of course, but I
> > felt that there was an unmet need for a Naive Bayesian (NB) classifier
> > implementation to be offered as well.
> >
> > An NB classifier could be bootstrapped up with partially labelled training
> > data as explained in the Nigam, McCallum, et al paper of 2000 "Text
> > Classification from Labeled and Unlabeled Documents using EM".
> >
> > So, if there isn't an NB code base out there already, I'd be happy to
> > contribute a very solid implementation that we've used in production for a
> > good 5 years.
> >
> > I'd have to adapt it to load the same training data format as the ME
> > classifier, but I guess that shouldn't be very difficult to do.
> >
> > I was wondering if there was some interest in adding an NB implementation
> > and I'd love to know who could I coordinate with if there is?
> >
> > Cohan Sujay Carlos
> > CEO, Aiaioo Labs, India
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
