hongwang600 / docred Goto Github PK

Thank you for your work!
I have questions about the SentModel mentioned in your paper.

1. Is the SentModel based on BiLstm or Bert?
2. I can't find the code about SentModel. Can you point it up?

Hi. I'm trying to run your code but am experiencing a bit of confusion. Looking at the BERT model (I believe in the BiLSTM module) I'm confused which part is the first phase and which is the second. It seems to me that there's just one large phase rather than two.

I've checked Issue #3 and it seems that the code for the two-step process is contained in the rel_exist_bert_cls_sep branch? However, I'm also a little confused as to where the first and second steps take place.

Could you provide some tips on where I should be looking and how I should be running the code properly? Thanks!

Edit

Reading the paper again, in "3.2 Implementation Details" you state in the second paragraph:

In the first step, we set the relation label for all relational instances to be 1, while the label for all N/A relations to be 0. We randomly sample N/A relations at a ratio 3:1 within a batch. In the second step, we train a new model using only relational instances, and the specific relation label is kept in this step.

I initially thought that you "pretrain" a model in the first step using binary classification and further fine-tune the model in the second step. However, if you train a new model in the second step, how is the information from the first step used?

Dear authors:
Thanks for your implementation with BERT on the DocRED. I have a question that the value of 'not NA acc' is quite large when training, and when the model converges, it even approaches 1. But the test F1 is more normal with a number about 0.54. Beyond that, I find that the value of original implementation (ACL-19) with LSTM seems in line with the final test F1. Thus I want to know why the 'not NA acc' and 'test F1' are so different in training.
Looking for your reply!

Dear authors, thanks for your efforts. I am planning to use your Bert implementation as a baseline for my MSc project concerning document-level RE. In the final part of the paper, you compare the performance of the sentence-encoding model and BiLSTM. Would you like to tell me if the BiLSTM refers to the baseline model in the DocRED paper?

Where is ' checkpoint_BiLSTM_bert_relation_exist_cls'? How to generate it?
Thanks