Closed hulianyuyy closed 2 years ago
I have witnessed the latest version of your paper where results are updated. This issue is to be closed.
However, i'd like to ask what it is to make you boost it from 22.1 to 21.2 compared to the last version of code? Thanks in advance.
@hulianyuyy We add ablations about the temporature and loss weight in supplementary, which can be found in Table 7 & 8 of our arXiv version.
May i also ask that the evaluation tool (python or scilte) will or won't affect the performance?
@hulianyuyy From my experience, the evaluation differences between these tools about WER are less than 0.1%.
Then how could i reproduce the results of iterative training as mentioned in tab.1 in the paper ?
@hulianyuyy You can follow their papers details and try to implement them, we didn't revised this part of codes. Besides, there still a performance gap between my implementation and their results and we didn't pay much attention on this.
Many thanks for your reply of your work which will benefit the community a lot.
I want to make sure that you report 22.1 Dev WER and 23.0 Test WER, while 21.2 Dev WER and 22.3 Test WER of released pretrained model ? Thanks in advance for response!