Addidng a layer of CRF on top of BERT + linear could improve performance and I think it's quite state of the art. I see it like a boolean option to add when creating the model.
I'd be happy to write a PR if this idea is supported and no volunteers show up :)
I think this would be an excellent addition. I had an implementation with a LSTM on top but it didn't seem to make a significant improvement, so I didn't push it. If a CRF improves performance, it would be most welcome.
An LSTM wouldn't improve because does kind of the same as BERT. That's why one the best previous architectures before LMs was BiLSTM+CRF
Let's see if someone volunteers for that :) otherwise I'll eventually do it
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.
Let's see if someone volunteers for that :) otherwise I'll eventually do it
@marctorsoc Any progress on that :) ? Thx
not really. If you wanna go with it, I'm very happy to help but can't get it on my stack as I'm really busy these weeks
Most helpful comment
Let's see if someone volunteers for that :) otherwise I'll eventually do it