Recovery Should Never Deviate from Ground Truth: Mitigating Exposure Bias in Neural Machine Translation

Jianfei He, Shichao Sun, Xiaohua Jia, Wenjie Li


Abstract
In Neural Machine Translation, models are often trained with teacher forcing and suffer from exposure bias due to the discrepancy between training and inference. Current token-level solutions, such as scheduled sampling, aim to maximize the model’s capability to recover from errors. Their loss functions have a side effect: a sequence with errors may have a larger probability than the ground truth. The consequence is that the generated sequences may recover too much and deviate from the ground truth. This side effect is verified in our experiments. To address this issue, we propose using token-level contrastive learning to coordinate three training objectives: the usual MLE objective, an objective for recovery from errors, and a new objective to explicitly constrain the recovery in a scope that does not impact the ground truth. Our empirical analysis shows that this method effectively achieves these objectives in training and reduces the frequency with which the third objective is violated. We conduct experiments on three language pairs: German-English, Russian-English, and English-Russian. Results show that our method outperforms the vanilla Transformer and other methods addressing the exposure bias.
Anthology ID:
2024.eamt-1.10
Volume:
Proceedings of the 25th Annual Conference of the European Association for Machine Translation (Volume 1)
Month:
June
Year:
2024
Address:
Sheffield, UK
Editors:
Carolina Scarton, Charlotte Prescott, Chris Bayliss, Chris Oakley, Joanna Wright, Stuart Wrigley, Xingyi Song, Edward Gow-Smith, Rachel Bawden, Víctor M Sánchez-Cartagena, Patrick Cadwell, Ekaterina Lapshinova-Koltunski, Vera Cabarrão, Konstantinos Chatzitheodorou, Mary Nurminen, Diptesh Kanojia, Helena Moniz
Venue:
EAMT
SIG:
Publisher:
European Association for Machine Translation (EAMT)
Note:
Pages:
68–79
Language:
URL:
https://aclanthology.org/2024.eamt-1.10
DOI:
Bibkey:
Cite (ACL):
Jianfei He, Shichao Sun, Xiaohua Jia, and Wenjie Li. 2024. Recovery Should Never Deviate from Ground Truth: Mitigating Exposure Bias in Neural Machine Translation. In Proceedings of the 25th Annual Conference of the European Association for Machine Translation (Volume 1), pages 68–79, Sheffield, UK. European Association for Machine Translation (EAMT).
Cite (Informal):
Recovery Should Never Deviate from Ground Truth: Mitigating Exposure Bias in Neural Machine Translation (He et al., EAMT 2024)
Copy Citation:
PDF:
https://aclanthology.org/2024.eamt-1.10.pdf