opennlp-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Olivier Grisel <olivier.gri...@ensta.org>
Subject Re: Distributing our statistical models
Date Wed, 19 Jan 2011 23:40:19 GMT
2011/1/20 Jörn Kottmann <kottmann@gmail.com>:
> On 1/19/11 11:43 PM, Jörn Kottmann wrote:
>>
>> On 1/19/11 11:29 PM, Olivier Grisel wrote:
>>>>
>>>> The models currently do not have a real version, but they contain
>>>> the opennlp tools version for compatibility checking. Anyway
>>>> adding such a version could be done easily.
>>>
>>> You just need to write a very short pom.xml file to build a versioned
>>> jar that holds the model (the .bin.gz file in folder
>>> src/main/resource/opennlp/ for instance). Then you can use mvn deploy
>>> to deploy the jar as usual to a maven repo server when doing the
>>> opennlp release.
>>>
>>
>> Does that mean, that OpenNLP has to load the model via the classpath,
>> instead from a file ?
>
> If it is one the classpath it would be really easy to open an InputStream
> pointing
> to the real model. That InputStream can then be used as usual to initialize
> a component.
>
> Sounds like a nice solution for all the people which are using some kind
> of dependency management and want to try out OpenNLP quickly.

Yes.

> For production deployments it seems not to be a nice solution, since it is
> more difficult to update the model while an application is running when it
> is loaded from
> the classpath.

It's useful even for production: very few people know about training
custom statistical NLP models and will go with the default. And those
who know about custom models will setup their application allow the
use of the default models from the classpath if none is configured in
an application specific configuration file or runtime setup such as
OSGi for instance.

-- 
Olivier
http://twitter.com/ogrisel - http://github.com/ogrisel

Mime
View raw message