search

salesforce / decaNLP

The Natural Language Decathlon: A Multitask Challenge for NLP
BSD 3-Clause "New" or "Revised" License
2.34k stars 474 forks source link

updating ELMo implementation #45

Closed Mehrad0711 closed 5 years ago

Mehrad0711 commented 5 years ago

Hi, I noticed there's been some work to have ELMo word representations for decaNLP. I have an implementation of it for my own work. Thought this PR would be helpful if others want to use it as well.

salesforce-cla[bot] commented 5 years ago

Thanks for the contribution! Before we can merge this, we need @Mehrad0711 to sign the Salesforce.com Contributor License Agreement.

bmccann commented 5 years ago

Thanks! I have elmo running locally, but I haven't yet seen improvements from using any subset of the output layers. Have you seen any improvements with your version?

Mehrad0711 commented 5 years ago

Hi, So I have only trained and tested the model on our own task (which is natural language to code translation) and unfortunately, it's doing better when not using ELMo. However, I'm still working on it to see if we can improve it.

bmccann commented 5 years ago

So odd; I really expected ELMo to help right out of the box. I've tried a few different ways of using it with no success, but I've pushed the most recent version for now.

Mehrad0711 commented 5 years ago

Same here! I was suspecting that appending GloVe to ELMo embeddings was downgrading its performance since GloVe is word based as opposed to character based. However, I see that you have tried both variations. I will close this PR then until we can get better results.