Generalization With Neuro-NER and overfiting disease

[Parvez-Khan-1]

Hello Frank, I used Neuro-ner to solve one of my problem , it working fine but looks like it suffered with over fitting disease of ML.It showing 99.85% on train_set , 100.00% on validation_set and 82.73% on test_set. I need your suggestion to get rid of this problem so that my model is a generalized model which will perform good on unseen data also, looking forward to hear from you.

Thanks ParvezKhan

[grafael]

Increasing dropout values may help you.

[Parvez-Khan-1]

If i increase dropout rate so, should i also need to increase learning rate?. Currently i have 0.5 drop out rate and 0.005 learning rate.

[grafael]

No, as far as I know, changing the learning rate will not prevent the overfitting.

[JohnGiorgi]

Hi ParvezKhan,

In order to address over-fitting, there is really two major things you can do: regularization (increase dropout) and get more data.

Is there anyway you can gather more data for your task? Or perhaps, use your trained model to tag some unlabeled text and bootstrap the process like that.