This perturbation translates randomly picked words in the text from English to other languages (e.g., German). It can be used to test the robustness of a model in a multilingual setting.
Author names:
- Genta Indra Winata ([email protected], The Hong Kong University of Science and Technology),
- Samuel Cahyawijaya ([email protected], The Hong Kong University of Science and Technology)
- Bryan Wilie ([email protected], Institut Teknologi Bandung).
This transformation acts as a perturbation to test robustness. Few words were picked at random with a probability and translated to the target language.
This perturbation would benefit all tasks with a sentence/paragraph/document as input like text classification, text generation, etc.
(1) Mixed-Language Training (Published in AAAI 2020)
@inproceedings{liu2020attention,
title={Attention-informed mixed-language training for zero-shot cross-lingual task-oriented dialogue systems},
author={Liu, Zihan and Winata, Genta Indra and Lin, Zhaojiang and Xu, Peng and Fung, Pascale},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={34},
number={05},
pages={8433--8440},
year={2020}
}(2) Continual Mixed-Language Pre-Training (Accepted in ACL Findings 2021)
@article{liu2021continual,
title={Continual Mixed-Language Pre-Training for Extremely Low-Resource Neural Machine Translation},
author={Liu, Zihan and Winata, Genta Indra and Fung, Pascale},
journal={arXiv preprint arXiv:2105.03953},
year={2021}
}
(3) M2M100 Model, Beyond English-Centric Multilingual Machine Translation (JMLR 2021)
@article{fan2021beyond,
title={Beyond english-centric multilingual machine translation},
author={Fan, Angela and Bhosale, Shruti and Schwenk, Holger and Ma, Zhiyi and El-Kishky, Ahmed and Goyal, Siddharth and Baines, Mandeep and Celebi, Onur and Wenzek, Guillaume and Chaudhary, Vishrav and others},
journal={Journal of Machine Learning Research},
volume={22},
number={107},
pages={1--48},
year={2021}
}
The transformation's outputs are dependent on the accuracy of the individual translation models and generally would generate simpler text or more popularly used text.