@inproceedings{wang-etal-2019-improving-back,
title = "Improving Back-Translation with Uncertainty-based Confidence Estimation",
author = "Wang, Shuo and
Liu, Yang and
Wang, Chao and
Luan, Huanbo and
Sun, Maosong",
booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
month = nov,
year = "2019",
address = "Hong Kong, China",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/D19-1073",
doi = "10.18653/v1/D19-1073",
pages = "791--802",
}
My literature review slide (in Japanese) [speakerdeck]
1. What is it?
In this paper, they proposed the method to qualify the confidence value for pseudo-data from back-translation.
They used (sentence-level | word-level) confidence score
sentence-level C: modifying the likelihood function during training forward model
word-level: word-level confidence cannot use like sentence-level, so they used as an attention weights.
2. What is amazing compared to previous studies?
They calculated the uncertainty score by using the confidence value.
3. Where is the key to technologies and techniques?
There are 3 types of uncertainty,
model: estimate the parameters inner the model
data: label unmatching, overlapping class, ...
distribution: the difference of distribution between the train and test data
They focused on model uncertainty using Monte Carlo dropout sampling.
4. How did validate it?
In MT, they calculated the BLEU score for each method of confidence value.
As above, the method CEV is the best,
Combination of Expectation and Variance
equation is below,
After that, they achieve the better the score of Back-Translation using uncertainty-based confidence method than the Neural Quality Estimation method.
5. Is there a discussion?
The big difference from QE is that their system does not need any other models.
a1da4
commented
5 years ago
0. Paper
@inproceedings{wang-etal-2019-improving-back, title = "Improving Back-Translation with Uncertainty-based Confidence Estimation", author = "Wang, Shuo and Liu, Yang and Wang, Chao and Luan, Huanbo and Sun, Maosong", booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)", month = nov, year = "2019", address = "Hong Kong, China", publisher = "Association for Computational Linguistics", url = "https://www.aclweb.org/anthology/D19-1073", doi = "10.18653/v1/D19-1073", pages = "791--802", } My literature review slide (in Japanese) [speakerdeck]
1. What is it?
In this paper, they proposed the method to qualify the confidence value for pseudo-data from back-translation. They used (sentence-level | word-level) confidence score
sentence-level C: modifying the likelihood function during training forward model
word-level: word-level confidence cannot use like sentence-level, so they used as an attention weights.
2. What is amazing compared to previous studies?
They calculated the uncertainty score by using the confidence value.
3. Where is the key to technologies and techniques?
There are 3 types of uncertainty,
They focused on model uncertainty using Monte Carlo dropout sampling.
4. How did validate it?
In MT, they calculated the BLEU score for each method of confidence value.
As above, the method CEV is the best,
After that, they achieve the better the score of Back-Translation using uncertainty-based confidence method than the Neural Quality Estimation method.
5. Is there a discussion?
The big difference from QE is that their system does not need any other models.
6. Which paper should read next?
Uncertainty based study