search

dwadden / dygiepp

Span-based system for named entity, relation, and event extraction.
MIT License
575 stars 120 forks source link

Hi, nice work! In event extraction, do you use BERT or Roberta? #67

Closed jianliu-ml closed 3 years ago

jianliu-ml commented 3 years ago

It seems that you set "roberta-base" in event.jsonnet

YeDeming commented 3 years ago

And the shape of word embedding in the downloaded ace-event model is (50265, 768). Do you use bert-large-cased or bert-large-uncased in the paper?

dwadden commented 3 years ago

You're right, the current master branch of the code uses RoBERTa rather than BERT, which is different from the paper. The code's been updated substantially since the paper was released; in particular, the configuration is much easier to deal with now. If you'd like to train the model using a different BERT, you should be able to change the relevant config. Let me know if you need help doing this.

If you need the code as it was for the paper, you can use this branch. From the ACE event config, it looks like I was using bert-large-cased.

I wouldn't recommend dealing with the older version of the code unless you need to for some reason.