[NLPL Task Force (A)] [NLPL Board] Hosting of ELMo models

[Stephan Oepen]

dear frankie,

thanks for making contact, and for the suggestion to inject these
models into the NLPL vectors repository!

i am taking this thread off the general ‘contact’ address, to spare
the project steering group further traffic.

the NLPL repository has been primarily designed and maintained by
andrey kutuzov (copied) and myself.  i have posted an offer to host
these models on the GitHub issue you mentioned, so i hope the original
creators will find this possibility attractive and get back to us
about one-time download, metadata, and such.

best wishes, oe

On Mon, Oct 8, 2018 at 6:49 PM Frankie Robertson
<frankie.r.robertson at student.jyu.fi> wrote:
>
> Dear NLPL project management,
>
> I have been trying to use a pretrained ELMo language model/word
> vectors in my work on word-sense disambiguation. They have been made
> available by researchers from the Harbin Institute of Technology at
> https://github.com/HIT-SCIR/ELMoForManyLangs . As per
> https://github.com/HIT-SCIR/ELMoForManyLangs/issues/5 -- there appears
> to be some troubles with web hosting. I'm writing to see if this is
> something NLPL might be able to help with, perhaps under the same
> umbrella of the existing word vectors which you have made available.
> As I understand these ELMo models are trained on the same CoNLL 17 raw
> tokenized text. I have only been able to obtain the model for Finnish
> successfully, so we would need to contact the authors to obtain the
> rest.
>
> I suppose that long term it may be better to have a clean path to ELMo
> word vectors based on the mainline AllenNLP code -- but perhaps simply
> rehosting HIT's resources could be useful for people in the short
> term?
>
> Regards,
> Frankie Robertson