opennlp-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Joern Kottmann <kottm...@gmail.com>
Subject Re: Joining the group
Date Wed, 05 Jul 2017 18:49:01 GMT
I spent some time on the coref component and updated it to now use
OpenNLP 1.6.0, it can also be trained on MUC 6. Not sure if the models
are any good, we need to work on evaluation for that.

We created an OpenNLP Improvement Proposal (NIP) to get it into a
better shape again.

The document is still empty, but it will be located here:
https://cwiki.apache.org/confluence/display/OPENNLP/NIP-3%3A+Revive+the+coreference+component

Jörn

On Thu, Jun 29, 2017 at 7:14 PM, Joern Kottmann <kottmann@gmail.com> wrote:
> Hello,
>
> there are a few problems we have with it. It would be very good if you
> can help us to solve those.
>
> Basically we would need to get it into the following state:
> - Have a data set it can be trained on
> - Implement evaluation for it
> - Write some documentation
>
> As far as I remember we somehow got stuck with getting it trained correctly.
>
> If we get it into a state where we can train a working model we can
> include it again in our main release.
>
> Jörn
>
>
>
>
> On Thu, Jun 29, 2017 at 7:09 PM, Ashkan Gholamzadeh
> <agholamzadeh@apple.com> wrote:
>> Hi,
>>
>> I have been using coreference resolution package in 1.5.3 recently and had some good
experience with it. My understanding is that it’s not supported anymore in recent version
of OpenNLP. I am wondering if I can join the group to work to put back coreference along with
new word net lib package back into  latest version as I think there are lots of people out
there that want to use it. I used it in a multi-threaded environment and there are some work
that need to be done to make it thread safe. It’s much faster than Stanford CoreNLP that
I used for coreference and accuracy is comparable. There are certain functionalities like
finding the most representative entity that can be added to current package to enhance its
current functionality. Training a new model would also be something that can be done to improve
the accuracy of algorithm.
>>
>> Please advice,
>>
>> Ashkan
>>

Mime
View raw message