Need help with NER cross validation using API

[Michail Drozdov <mi...@gmail.com>]

Hello,

I want to build NER application for lithuanian language, so I'm trying
different configurations of feature generators. When tied to evaluate
resulting models I encountered following problem: when training models I
build feature generators using API (not descriptor files), so have many
AdaptiveFeatureGenerator instances, but TokenNameFinderCrossValidator
constructor can only take byte[] featureGeneratorBytes.

So my question is, how can pass my future generator instances to
TokenNameFinderCrossValidator constructor?

I'm new to java and opennlp so perhaps I miss something obvious, but I
can't find a way.

Thank you in advance,
Michail Drozdov

Re: Need help with NER cross validation using API

[Jörn Kottmann <ko...@gmail.com>]

Hello,

a custom feature generation is often defined using XML. This XML file
should be read into a byte array and then passed to the Token Name 
Finder Cross Validator.
Thats why this constructor is there.

In case a user needs more control the only other option is to implement 
a Token Name Finder Factory
and pass it in. The Factory class can create the feature generator.

I am not sure if that was added already for the 1.5.3 release or is 
trunk only. Anyway I suggest to use the latest
trunk version. It has some new/fixed classifiers which increase the 
performance of the name finder.

Jörn


On 08/26/2014 10:48 AM, Michail Drozdov wrote:
> Hello,
>
> I want to build NER application for lithuanian language, so I'm trying
> different configurations of feature generators. When tied to evaluate
> resulting models I encountered following problem: when training models I
> build feature generators using API (not descriptor files), so have many
> AdaptiveFeatureGenerator instances, but TokenNameFinderCrossValidator
> constructor can only take byte[] featureGeneratorBytes.
>
> So my question is, how can pass my future generator instances to
> TokenNameFinderCrossValidator constructor?
>
> I'm new to java and opennlp so perhaps I miss something obvious, but I
> can't find a way.
>
> Thank you in advance,
> Michail Drozdov
>