Uncertainty-Aware Curriculum Learning for Neural Machine Translation
Requirements
- Python version >= 3.7.4
- Pytorch version >= 1.2.0
Getting Started
Data Preprocessing
We use the standard validation and test sets provided in each translation task.
Language Models
The main language models used in this paper can be obtained from the following links.
Data Uncertainty
Follow these steps to calculate the data uncertainty and generate the data difficulty JSON file:
1.Calculate the perplexity of each sentence.
2.Sort from low to high.
3.According to your experimental requirements, divide the sorted data set into several bins.
4.Extract 1000 pairs of sentences from each bin to verify the model uncertainty for a certain stage.
5.Build data difficulty JSON file. This is an example of data difficulty JSON file for 4 bins. These numbers are the indices of the sentences in a data file, the train_set represents all the data of the training set, the esti_set represents the estimation set.
{
"train_set": [
[116518, 41568, 13049, ..., 39342, 23659, 76413],
[12051, 113004, 57498, ..., 51064, 47300, 47552],
[73186, 50806, 17741, ..., 94891, 55986, 44589],
[69885, 114662, 32893, ..., 103985, 85597, 84899]
],
"esti_set": [
[28948, 87465, 7934, ..., 7839, 89179, 55998],
[297, 84844, 4712, ..., 112400, 105640, 47525],
[115014, 71806, 46151, ..., 41996, 43563, 95774],
[22106, 66255, 72142, ..., 16703, 45681, 5157]
]
}
Training
An example of a training script could be found in the script folder. Most parameters are quite obvious. Some parameters need to be specially set are explained as follows：
--file_prefix Specify the directory of dataset.
--difficulty_json Specify the path of data difficulty JSON file.
--fold_name Specify the directory to store models.
Translation
An example of a translation script could be found in the script folder.