ontonote4 checkpoint and experimental replication problems

[founting]

Hi, I met some problems when replicating your experiments.

1. I use the ontonote4 checkpint you provided to do prediction. It shows that the crf.transitions has a shape torch.Size([20, 20]) from checkpoint, but actually this dataset only has 17 labels and the crf.transitions should be [19, 19]. So could you please check the checkpoint file of ontonote4?
2. I used a single GPU to train the models on weibo and ontonote4 dataset without changing any code or paramters. However the best F1 score of weibo is 0.68 and the ontonote4 is 0.80, which is lower than your result. If it is because you used the ditributed training, would you please provide the detailed paramters of distributed training, or the training parameters of single GPU that could reach your scores?

Many thanks in advance.

[liuwei1206]

Hi,

I share the label.txt for note4 with you and you can try again.

" X O B-PER M-PER E-PER S-PER B-LOC M-LOC E-LOC S-LOC B-GPE M-GPE E-GPE S-GPE B-ORG M-ORG E-ORG S-ORG "

As for the distributed training, I am sorry to say I forget the detailed parameters. Because it was a long time ago and I have resigned from the company.

[founting]

Hi,

I share the label.txt for note4 with you and you can try again.

" X O B-PER M-PER E-PER S-PER B-LOC M-LOC E-LOC S-LOC B-GPE M-GPE E-GPE S-GPE B-ORG M-ORG E-ORG S-ORG "

As for the distributed training, I am sorry to say I forget the detailed parameters. Because it was a long time ago and I have resigned from the company.

Thank you for your prompt reply! But what is the label 'X' used for ? It does not appear in the ontonote4 dataset.

[liuwei1206]

Hi,

I think it is just an empty label, we don't use it label data.

[liuwei1206]

This may cause why the transition shape is [20, 20] instead of [19, 19]