@inproceedings{he-etal-2024-recovery,
title = "Recovery Should Never Deviate from Ground Truth: Mitigating Exposure Bias in Neural Machine Translation",
author = "He, Jianfei and
Sun, Shichao and
Jia, Xiaohua and
Li, Wenjie",
editor = "Scarton, Carolina and
Prescott, Charlotte and
Bayliss, Chris and
Oakley, Chris and
Wright, Joanna and
Wrigley, Stuart and
Song, Xingyi and
Gow-Smith, Edward and
Bawden, Rachel and
S{\'a}nchez-Cartagena, V{\'\i}ctor M and
Cadwell, Patrick and
Lapshinova-Koltunski, Ekaterina and
Cabarr{\~a}o, Vera and
Chatzitheodorou, Konstantinos and
Nurminen, Mary and
Kanojia, Diptesh and
Moniz, Helena",
booktitle = "Proceedings of the 25th Annual Conference of the European Association for Machine Translation (Volume 1)",
month = jun,
year = "2024",
address = "Sheffield, UK",
publisher = "European Association for Machine Translation (EAMT)",
url = "https://aclanthology.org/2024.eamt-1.10",
pages = "68--79",
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.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="he-etal-2024-recovery">
<titleInfo>
<title>Recovery Should Never Deviate from Ground Truth: Mitigating Exposure Bias in Neural Machine Translation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jianfei</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shichao</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaohua</namePart>
<namePart type="family">Jia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenjie</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-06</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 25th Annual Conference of the European Association for Machine Translation (Volume 1)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Carolina</namePart>
<namePart type="family">Scarton</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Charlotte</namePart>
<namePart type="family">Prescott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chris</namePart>
<namePart type="family">Bayliss</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chris</namePart>
<namePart type="family">Oakley</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joanna</namePart>
<namePart type="family">Wright</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Stuart</namePart>
<namePart type="family">Wrigley</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xingyi</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Edward</namePart>
<namePart type="family">Gow-Smith</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rachel</namePart>
<namePart type="family">Bawden</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Víctor</namePart>
<namePart type="given">M</namePart>
<namePart type="family">Sánchez-Cartagena</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Patrick</namePart>
<namePart type="family">Cadwell</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ekaterina</namePart>
<namePart type="family">Lapshinova-Koltunski</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vera</namePart>
<namePart type="family">Cabarrão</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Konstantinos</namePart>
<namePart type="family">Chatzitheodorou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mary</namePart>
<namePart type="family">Nurminen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Diptesh</namePart>
<namePart type="family">Kanojia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Helena</namePart>
<namePart type="family">Moniz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>European Association for Machine Translation (EAMT)</publisher>
<place>
<placeTerm type="text">Sheffield, UK</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<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.</abstract>
<identifier type="citekey">he-etal-2024-recovery</identifier>
<location>
<url>https://aclanthology.org/2024.eamt-1.10</url>
</location>
<part>
<date>2024-06</date>
<extent unit="page">
<start>68</start>
<end>79</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Recovery Should Never Deviate from Ground Truth: Mitigating Exposure Bias in Neural Machine Translation
%A He, Jianfei
%A Sun, Shichao
%A Jia, Xiaohua
%A Li, Wenjie
%Y Scarton, Carolina
%Y Prescott, Charlotte
%Y Bayliss, Chris
%Y Oakley, Chris
%Y Wright, Joanna
%Y Wrigley, Stuart
%Y Song, Xingyi
%Y Gow-Smith, Edward
%Y Bawden, Rachel
%Y Sánchez-Cartagena, Víctor M.
%Y Cadwell, Patrick
%Y Lapshinova-Koltunski, Ekaterina
%Y Cabarrão, Vera
%Y Chatzitheodorou, Konstantinos
%Y Nurminen, Mary
%Y Kanojia, Diptesh
%Y Moniz, Helena
%S Proceedings of the 25th Annual Conference of the European Association for Machine Translation (Volume 1)
%D 2024
%8 June
%I European Association for Machine Translation (EAMT)
%C Sheffield, UK
%F he-etal-2024-recovery
%X 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.
%U https://aclanthology.org/2024.eamt-1.10
%P 68-79
Markdown (Informal)
[Recovery Should Never Deviate from Ground Truth: Mitigating Exposure Bias in Neural Machine Translation](https://aclanthology.org/2024.eamt-1.10) (He et al., EAMT 2024)
ACL