[QUESTION] What type of relation extraction is SpanBERT suited for? about spanbert HOT 3 CLOSED

Interesting question! It is admittedly hard to say for certain what will happen in your case. But, I'd like to think--after controlling for other pretraining details--SpanBERT is still more used to longer spans than comparable models. If and when you get results, I'd love to hear about them. Thanks!

from spanbert.

Interesting question! It is admittedly hard to say for certain what will happen in your case. But, I'd like to think--after controlling for other pretraining details--SpanBERT is still more used to longer spans than comparable models. If and when you get results, I'd love to hear about them. Thanks!

I'm about to test the model, just a quick question: a data point in TACRED has the following format:

Considering the fact that for longer spans, it may not make much sense to have NER tags, should I provide the NER tags when I'm converting my inputs to TACRED format? Or is it enough just to include spans and their start and end indices in the token list? I'm writing a new run_*.py class for my task, but I'm basically trying to figure out which elements in a record are essential to be fed to SpanBERT for the type of relation extraction I described earlier.

from spanbert.

I think it should be ok to skip the tags for your task.

from spanbert.