NLP classification tasks often use F1 scores for evaluation, but they use cross entropy for optimization, and if using unbalanced data, there exists large gap between them. Therefore, they proposed a loss of DSC that can be interpreted as smoothing out F1 and multiplied by a factor that reduces the coefficient of easily classifiable samples to zero. They confirmed that the loss was effective across a number of models and datasets.
TL;DR
NLP classification tasks often use F1 scores for evaluation, but they use cross entropy for optimization, and if using unbalanced data, there exists large gap between them. Therefore, they proposed a loss of DSC that can be interpreted as smoothing out F1 and multiplied by a factor that reduces the coefficient of easily classifiable samples to zero. They confirmed that the loss was effective across a number of models and datasets.
Why it matters:
Paper URL
https://arxiv.org/abs/1911.02855
Submission Dates(yyyy/mm/dd)
2019/11/07
Authors and institutions
Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, Jiwei Li
Methods
Results
Comments
ACL2020