Some question about the value of action product by policy network about dsner-pytorch HOT 2 CLOSED

The coder is the PyTorch implementation of the following repository: https://github.com/rainarch/DSNER
The issue that you mention is already addressed in the following paper and also proposed a solution:
https://www.aclweb.org/anthology/D19-6125/

from dsner-pytorch.

Hello, I have read the article you mentioned in the comment section, in which you used the difference of F score as a reward, which is very interesting. But I tried to repeat your experiment and found several problems. The first problem is that the article mentions that the NER model should be pre-trained before retraining, but you don't seem to have given a specific number of pre-training rounds in the article. The second problem is that when we update the selector with the difference of F score as a reward, the reward value is oscillating and doesn't converge, which means that the selector isn't picking the instance very well, thus leading to the poor performance of the NER model. I would appreciate it if you could give me some advice.

from dsner-pytorch.