What if i would like to use Electra model from huggingface transformers? about bond

[cliang1453]

Hi,

I think you mainly need to edit convert_examples_to_features in data_utils.py to fit the Electra input pattern, and edit modeling_roberta.py.

[marcusau]

Sure, Thanks a lot.

Also , my understanding is . first run the stage 1 by run_ner.py , and then on the top of the stage 1 output to run self-learning.py.

Please correct me if i am wrong.

Thanks a lot.

Marcus

[cliang1453]

Hi Marcus, sorry for the confusion. Running run_self_training_ner.py will directly produce the two-stage results. You don't need to run it on top of the stage 1 output.

[marcusau]

got it .

Thanks a lot.

What i just wanna know is ...... If i dont have the hp label in my dataset. how to produce it ??

Thanks a lot :)

[cliang1453]

Hi Marcus, we produced hp labels for CoNLL03 using pos tagging and rule-based method, and see little improvement unfortunately. Hence, we decided not to use hp labels for training, by not setting args.self_training_hp_label to be 2 or 3 (

BOND/model_utils.py

Line 48 in 3651a92

elif args.self_training_hp_label == 2:

). We plan to remove these redundant codes that cause confusion.

[marcusau]

thanks..

If hp label is not so useful in BOND, I can run the Bond training by run_self_training_ner.py and set args.self_training_hp_label not to be 2 or 3 in the line 570 of run_self_training_ner.py.
parser.add_argument('--self_training_hp_label', type = float, default = 0, help = 'use high precision label.')

Thanks for that 👍

Marcus